Browse every tracked release and narrow the catalog by capabilities, openness, modalities, token limits, pricing, cost per task, and benchmark results.
SpaceXAI's lower-cost video model on the Imagine API, listed as grok-imagine-video-1.5-lite. It accepts text and image input and returns video. Listed rates are $0.02 per second at 480p, $0.03 at 720p, and $0.14 at 1080p, plus $0.01 per input image. The model page headline of $0.020 per second matches the 480p default, one quarter of [Grok Imagine Video 1.5](/xai/grok-imagine-video/1-5) at $0.08. The Lite registry entry does not include the audio input modality used for preset voices on 1.5. The Batch API is supported, at 10 requests per second, in us-east-1 and us-west-2.
Google's frontier Gemini generation, rolling out first to trusted cyber defenders through the Fairwind Program. Scores 77.9% on DeepSWE v1.1 and 68.9% on the Vals Index, with a 1M-token context window and a 1M-token output limit, up from 64K. Introductory API price is $2/M input and $10/M output, with cached input at $0.10/M. A measured price per task at high effort is $1.99. After the introductory period the list price moves to $4/$20. Paid API, AI Studio, and consumer access are still ahead, and Google has not published an API model ID. Comparable frontier models: [GPT Astra 6](/openai/gpt-astra/6), [Claude Opus 5.5](/anthropic/opus/5-5), and [Claude Fable 5.1](/anthropic/fable/5-1).
GPT-6.1 generation of OpenAI's Sol tier: near-[GPT Astra 6](/openai/gpt-astra/6) performance for coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices. Standard rates match [GPT Sol 6](/openai/gpt-sol/6) at $2/M input and $10/M output. Cache reads fall from $0.20/M to $0.10/M (95% off uncached input) and cache writes stay $2.50/M. A measured price per task at max effort is $0.72, below Sol 6 at $1.06. Available as gpt-6.1-sol in the API, ChatGPT Work, and Codex for Plus, Pro, Business, Enterprise, and Edu. Not yet available in Chat. 1.05M context, up to 922k input, 128k max output, knowledge cutoff April 30, 2026. Reasoning effort is low, medium (the default), high, xhigh, or max. Fast mode is 2x Standard and is unavailable with EU data residency. Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the full request. Uses the same safeguards stack as Astra.
Second model in the Claude 5.5 family, and a faster complement to [Claude Opus 5.5](/anthropic/opus/5-5) for well-scoped everyday tasks, bug fixes, and polished documents, slides, and spreadsheets. Scores 70.6% on Terminal-Bench 4.0, versus 10.3% for [Claude Sonnet 5](/anthropic/sonnet/5) and 66.4% for Opus 5.5 at xhigh. CursorBench 4.0 is 55.5% (Opus 5.5 is 57.8%). GDPval-AA v2.1 is 1844 Elo, two points below Opus 5.5. FrontierCode v1.1 Main is 52.1% at xhigh and 46.2% at max. List price stays $2/M input and $10/M output, with $0.20/M cache reads and $2.50/M cache writes. Anthropic reports up to 30% lower cost per task than Sonnet 5 and output that is more than 30% faster. A measured price per task is $7.60 at max effort and $1.08 at high effort, the Claude Platform default. Apps default to medium. Available as claude-sonnet-5-5 on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. 1M-token context, 128k max output, knowledge cutoff June 2026. First Sonnet with cyber safeguards and reasoning-extraction classifiers comparable to Opus 5.5. Thinking cannot be turned fully off: use between_tools to keep up-front thinking off. Claude Haiku 5.5 is still scheduled for the coming weeks.
MiniMax released M3.1 Flash Preview in MiniMax Code on September 27, 2026. The company describes it as a text model for everyday development: quick bug fixes and full feature work, with an emphasis on speed and stability. It is selectable beside [MiniMax M3](/minimax/minimax-m3/m3). MiniMax has not published a model card, benchmark table, list price, public API model ID, or weights. The models.dev index lists a 1M-token context and up to 512K output.
High-throughput serving ID for [Qwen 3.8-Max](/alibaba/qwen/3-8-max). Same 2.4T-parameter (95B active) weights, 1M-token context, and text, image, and video input. Alibaba Cloud international lists qwen3.8-max-prime at $4/M input, $0.50/M cache reads, and $12/M output, twice the $2/$6 rate recorded here for [Qwen 3.8-Max](/alibaba/qwen/3-8-max). Published Max scores carry over, including 67.7 on SWE-bench Pro and 86.6 on Terminal-Bench 2.1.
Google's flagship creative text-to-speech model, API ID gemini-3.8-flash-tts. Natural-language voice design, a library of 2,000+ voices, consented voice replication from a 30-second sample, and line-level style control across 130 languages. Google reports 71.4 on Hume AI's Voice Design Benchmark and 60.8 on accent modeling, and first place on Hume's Overall Quality Index. Input limit 8,192 tokens and output limit 16,384. Google AI Studio lists $0.50/M text input and $9/M audio output. Every clip carries a SynthID watermark. Rolling out in the Gemini API, Google AI Studio, and Google Vids. Volume sibling: [Gemini TTS 3.8 Flash-Lite](/google/gemini-tts/3-8-flash-lite).
High-volume sibling of [Gemini TTS 3.8 Flash](/google/gemini-tts/3-8-flash), API ID gemini-3.8-flash-lite-tts. Google positions it as the replacement for gemini-3.1-flash-tts-preview: same speech schema, 101 languages, and second place on Hume's Overall Quality Index. Input limit 8,192 tokens and output limit 16,384. Google AI Studio lists $0.50/M text input and $6/M audio output. Built for dubbing, read-aloud, and voice-agent cascades.
First model in Anthropic's Claude 5.5 family. Performs at the level of [Claude Fable 5.1](/anthropic/fable/5-1) on most work and costs about 40% less to run than [Claude Opus 5](/anthropic/opus/5) on typical workloads. Leads the announcement table on Terminal-Bench 4.0 (66.4% at xhigh vs 55.8% for Fable 5.1), FrontierCode v1.1 Main (54.4%), CursorBench 4.0 (57.8%), and GDPval-AA v2.1 (1846 Elo). Input and output tokens are $4 and $20 per million, 20% below Opus 5; cache reads are $0.20 per million, 60% below Opus 5. Output is more than 30% faster. Fast mode is available at 2.5x speed for $8/$40. Ships with Fable-class cyber, biology, and distillation safeguards, including preserved thinking. Available as claude-opus-5-5 on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. 1M-token context, 128k max output, knowledge cutoff June 2026. [Claude Sonnet 5.5](/anthropic/sonnet/5-5) followed on September 28. Claude Haiku 5.5 is still scheduled for the coming weeks.
GPT-6 generation of OpenAI's Sol tier for complex coding and agentic workflows, trained with similar methods as [GPT Astra 6](/openai/gpt-astra/6) at half the [GPT Sol 5.6](/openai/gpt-sol/5-6) promotional API price. Standard rates are $2/M input and $10/M output, with $0.20/M cache reads and $2.50/M cache writes. On AutomationBench, Sol at xhigh scores 33.2% at $0.27 per task, ahead of [Claude Opus 5](/anthropic/opus/5) at max (26.9%) at about 9% of that cost. Agents' Last Exam at max is 56.4%. DeepSWE v1.1 at max is 68.8%, within 1.1 points of [Claude Fable 5](/anthropic/fable/5) at xhigh. OSWorld 2.0 offline partial at xhigh is 60.5%. Available as gpt-6-sol in the API, ChatGPT Work, and Codex for Plus, Pro, Business, Enterprise, and Edu. 1.05M context, 128k max output, knowledge cutoff April 20, 2026. Fast mode is 2x Standard rates. Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output.
GPT-6 generation of OpenAI's volume Luna tier for focused, high-volume tasks such as summarizing, extracting, and answering quick questions. Standard API pricing is $0.10/M input and $0.50/M output, 50% below [GPT Luna 5.6](/openai/gpt-luna/5-6) promotional rates, with $0.01/M cache reads and $0.125/M cache writes. DeepSWE v1.1 at max scores 66.6%, comparable to [Claude Opus 5](/anthropic/opus/5) and [Claude Fable 5](/anthropic/fable/5) at medium effort at 93% and 96% lower cost per task. On AutomationBench, high-effort Luna improves 5.4 points over its predecessor at 58% lower cost per task. Available as gpt-6-luna in the API; ChatGPT Work and Codex for paid plans; Free and Go users can use it in the desktop app. Not yet in Chat at launch. 1.05M context, 128k max output, knowledge cutoff May 18, 2026. Fast mode is 2x Standard rates.
Xiaomi's open-weight omnimodal flagship: a 1.02T sparse MoE (42B active) with text, image, video, and audio input, a 1M-token context, and MIT weights as MiMo-V2.6-Pro-RL. Xiaomi's model card scores 71.9 on DeepSWE v1.1, 53.1 on AutomationBench v1.0.6, 34.9 on Terminal-Bench 4.0, and 89.9 on Terminal-Bench 2.1. That same table lists [Claude Opus 5](/anthropic/opus/5) at 74.0 / 49.0, [GPT Sol 5.6](/openai/gpt-sol/5-6) at 73.0 / 39.9, and [Claude Fable 5](/anthropic/fable/5) at 70.0 / 42.4 on DeepSWE v1.1 and Terminal-Bench 4.0. Overseas API ID mimo-v2.6-pro is $0.435/M uncached input, $0.0036/M cache hits, and $0.87/M output. Cache writes are free for a limited time. Hosted UltraSpeed (mimo-v2.6-pro-ultraspeed) is the same checkpoint at $4.35/$8.70. Measured price per task is $0.13. Weights are on Hugging Face; the API is on Xiaomi MiMo, AI Studio, MiMo Code, and MiMo Desktop.
Efficiency checkpoint of [MiMo V2.6 Pro](/xiaomi/mimo/v2-6-pro): 309B total / 15B active, the same omnimodal inputs and 1M-token context, MIT weights as MiMo-V2.6-Flash-RL. Xiaomi scores Flash at 67.9 on DeepSWE v1.1, 52.3 on AutomationBench v1.0.6, 28.8 on Terminal-Bench 4.0, and 87.6 on Terminal-Bench 2.1. Overseas API ID mimo-v2.6-flash is $0.14/M uncached input, $0.0028/M cache hits, and $0.28/M output. Batch inference is half of those uncached rates.
SpaceXAI's most capable model for coding and knowledge work, succeeding [Grok 4.6](/xai/grok/4-6) at the same $2/$6 per million token list price. A new, larger base with a longer RL run on multi-hour tasks scores 46.3% on CursorBench 4.0 (xhigh) versus 40.4% for 4.6, 38.0% on Terminal-Bench 4.0 versus 20.3%, and 71.0% on DeepSWE v1.1 at high effort. It leads the announcement table on EEBench (64.0%) and Harvey Legal Agent (19.6%) versus [GPT Sol 5.6](/openai/gpt-sol/5-6) and [Claude Fable 5.1](/anthropic/fable/5-1), and sits close to Fable 5.1 on AA Briefcase v1.1 (1,657 vs 1,678) and GDPval (1,695 vs 1,735). The EEBench bar chart in the same post puts [GPT Astra 6](/openai/gpt-astra/6) ahead at 69.3%. Available today in Cursor, Grok Build, the Grok API as grok-4.7, and third-party harnesses. Fast doubles output speed at 2x the price. 500K context, knowledge cutoff May 2026.
Alibaba's native omnimodal agent model for text, image, audio, and video input with text output. 1M-token context, function calling, web search, and default xhigh reasoning on Chat Completions and Responses as qwen3.8-omni-flash. Qwen reports 71.0 on WildClawBench-MM, 69.6 on UniClawBench, 63.3 on SWE-bench Pro, and 57.8 on DeepSWE v1.1, with audiovisual scores close to [Gemini 3.8 Flash](/google/gemini/3-8-flash). International Model Studio rates are $0.15/M input, $0.016/M cache-hit input, and $0.47/M output. Companion open-source Qwen-MM-Plugins cover long-form audio-visual workflows. Sibling of [Qwen 3.8-Max](/alibaba/qwen/3-8-max) and [Qwen 3.8-Flash-Next](/alibaba/qwen/3-8-flash-next).
Live WebSocket and WebRTC sibling of [Qwen Omni 3.8 Flash](/alibaba/qwen-omni/3-8-flash) for continuous audio-visual conversation. API ID qwen3.8-omni-flash-realtime streams camera and microphone input and returns text plus audio, with Skills, tool calls, spatial audio, and speech generation in 29 languages. Qwen reports about 83-85 text tokens/s and sub-second time to first audio packet on 6-20 second audio and audio-visual clips. Ships with open-source Qwen-Live Harness. Comparable live-dialogue products: [Gemini Live 3.8](/google/gemini-live/3-8) and [GPT Live 1](/openai/gpt-live/1).
Google's default low-latency native speech-to-speech model for real-time voice agents, succeeding gemini-3.1-flash-live-preview. Processes audio, image, video, and text with a 131k-token input window and 65k output. Supports asynchronous function calling, near real-time visual grounding, and mid-conversation switching across 97 languages. Google reports a second-place Speech Agent Arena ranking. API ID gemini-3.8-live on the Gemini Live API. Audio list rates $3/M input ($0.005/min) and $12/M output ($0.018/min); text is $0.75/$4.50. Rolling out in the Gemini API, Google AI Studio, Gemini Enterprise private preview, and Search Live. Comparable live-dialogue product: [GPT Live 1](/openai/gpt-live/1).
High-reasoning sibling of [Gemini Live 3.8](/google/gemini-live/3-8) for complex multi-step voice agents. Reasons and speaks at the same time, with configurable thinking_level (low, medium, or high) and async-only tool calls. Google reports #1 on the Speech to Speech Quality Index (82.6), 68.6% on tau-Voice, 35.1% on Sierra tau-Voice-banking, and 97.7% on Big Bench Audio. API ID gemini-3.8-live-extended-thinking. Same audio list rates as 3.8 Live. Rolling out in the Gemini API, Google AI Studio, Gemini Live, and Workspace Docs, Gmail, and Keep.
Moonshot AI's Kimi Code coding model, rolled out in place on the existing kimi-for-coding API ID. Official docs place overall coding and agent performance close to [Kimi K3](/moonshot/kimi/k3), with more efficient thinking than K2.7 Code. Supports the same reasoning_effort levels as K3 (low / high / max, default max). Image and video input, plus a 1M-token context window for all membership tiers. Thinking-off requests to K3 and K2.8 Preview are served by K2.8 Preview without thinking. Available in Kimi Code CLI, VS Code, and third-party tools via the Kimi Code API. Billed through Kimi Code membership quota rather than a published per-token card.
DeepSeek's new-architecture Flash model with native image understanding, a 552B-parameter backbone, and a 1M-token context. Causal Encoder-Decoder serving activates 8B parameters in prefill and 16B during decode. Official scores include 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, matching or beating [Claude Opus 5](/anthropic/opus/5) and [GPT Sol 5.6](/openai/gpt-sol/5-6) on those agent tasks. Available on the API as deepseek-flash, with MIT-licensed weights on Hugging Face.
Cohere's first North-family translation model is a 218B-total, 25B-active open MoE for 50+ languages. It scores 83.60 on WMT26 across all languages, or 84.36 with an agentic multi-pass workflow, ahead of DeepL NextGen (81.37), [Gemma 4](/google/gemma/4) 31B (79.46), [Command A+](/cohere/command-a/a-plus) (76.53), [GLM 5.2](/zhipu/glm/5-2) FP8 (76.50), and Google Translate (68.20). 16K input and 16K output. CC BY-NC 4.0 weights on Hugging Face; commercial access through Model Vault and RWS Language Weaver.
Cognition's most advanced coding model, post-trained from [Kimi K3](/moonshot/kimi/k3) (2.8T total parameters) with a single RL run across reasoning-effort levels. Scores 50.0% on FrontierCode 1.1 Main (within one point of [Claude Fable 5.1](/anthropic/fable/5-1) at a claimed 64% lower cost), 73.0% on DeepSWE v1.1, 92.8% on Terminal-Bench 2.1, and 27.3% on Terminal-Bench 4. Available in Devin Desktop and CLI at launch, with Devin Web and Fusion rolling out. No public standalone API or weights.
Default Images API model in the ChatGPT Images 2.5 generation: faster, higher-quality everyday image generation and editing from text and image inputs. OpenAI reports up to 50% lower latency than GPT-Image-2 while preserving subjects across reference-led edits, multi-turn refinements, and complex layouts. Available in ChatGPT, ChatGPT Work, and Codex as Images 2.5, and in the API as gpt-image-2.5-flare. Text input $5/M ($1.25/M cached), image input $8/M ($2/M cached), image output $30/M.
Premium Images API sibling of [GPT Image 2.5 Flare](/openai/gpt-image/2-5-flare) for editing-precision workflows such as campaign creative and product imagery. Same ChatGPT Images 2.5 generation, with longer generation times and tighter control across edits. API ID gpt-image-2.5-sunburst. Same listed token rates as Flare: $5/M text input, $8/M image input, $30/M image output.
Microsoft AI's strongest image model yet, in public preview on Microsoft Foundry with [MAI Image 2.6 Flash](/microsoft/mai-image/2-6-flash) as the production-speed sibling. Both add multi-image reference editing, web grounding, and dynamic aspect ratios up to 1.5K. Microsoft reports No. 2 on Arena for text-to-image and image editing as of 2026-09-04. Azure list rates: $5/M text input, $8/M image input, and $38/M image output. Available in MAI Playground and Foundry.
Lower-latency sibling of [MAI Image 2.6](/microsoft/mai-image/2-6) for high-throughput production. Microsoft says it generates images 2.8x faster than GPT-Image-2-Medium with 72% greater efficiency, at less than half the flagship list price. Azure rates: $1.75/M text input, $2.50/M image input, and $19/M image output. Same multi-reference editing, web grounding, and dynamic aspect ratios as 2.6.
Microsoft AI's most capable speech-to-text model, in public preview on Microsoft Foundry, Azure Speech, MAI Playground, and OpenRouter. Adds speaker diarization, word-level timestamps, clean vs verbatim styles, code switching, and automatic language detection across 60 languages. Microsoft reports 5.2% average WER on FLEURS across 60 languages (3.4% on the top 25) versus 3.7% top-25 WER and 43 languages for [MAI Transcribe 1.5](/microsoft/mai-transcribe/1-5). Beats [Gemini 3.5 Transcribe](/google/gemini-transcribe/3-5), GPT-Transcribe, Whisper V3-Large, and Scribe v2 on cited evals. Launch price $0.10 per hour of audio through 2026-12-31 (1.5 is $0.36/hour). Microsoft cites 10x batch speed vs GPT-Transcribe, 7x vs Scribe v2, and 5x vs Gemini 3.5 Transcribe.
OpenAI's most intelligent and aligned model yet, and the first GPT-6 generation flagship. State-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work. Saturates ARC-AGI-3 (99.9%) and ExploitBench (100%), scores 57.9% on Terminal-Bench 4.0 versus 37.3% for [GPT Sol 5.6](/openai/gpt-sol/5-6) and 55.8% for [Claude Fable 5.1](/anthropic/fable/5-1), and 64.6% on Terminal-Bench Science 0.1 versus 52.6% for Fable 5.1. Meets the Critical cybersecurity threshold under OpenAI's Preparedness Framework; shipping Astra refuses advanced exploit work, with less restrictive defensive access planned through OpenAI Daybreak. Rolling out first to a limited set of organizations, then ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API and AWS. API ID gpt-6-astra. Standard API pricing $10/M input and $50/M output, with $1/M cache reads and $12.50/M cache writes. Fast mode at 2x speed and 2x price. 1.05M context, 128k max output, knowledge cutoff April 30, 2026. Pro, Business, and Enterprise users also get GPT-6 Astra Pro.
Google's most intelligent Flash workhorse yet, released three weeks after [Gemini 3.7 Flash](/google/gemini/3-7-flash). Same introductory API price ($0.75/M input, $3.75/M output through 2026-12-31) but a higher price per completed task because the model works harder: $0.58 at high reasoning, up from $0.40 on 3.7 Flash. Scores 54.9% on HLE-Verified, 73.7% on DeepSWE v1.1, and 89.4% on Terminal-Bench 2.1. Available in the Gemini API as gemini-3.8-flash.
Cybersecurity-specialized sibling of [Gemini 3.8 Flash](/google/gemini/3-8-flash), available only to trusted defenders through Google's Fairwind Program. Frontier-level vulnerability discovery on CyberGym (86.2% Pass@1) and 71.0% on Google's internal 20-language benchmark, with CWE-Bench Pass@1 of 47.2% on the cost-accuracy Pareto frontier. No public API or listed token price.
Meta Superintelligence Labs' latest Muse Spark model for long-horizon agentic work, multitasking in a single thread, and coding. Available today in Muse Code and the Meta Model API with max reasoning. Meta reports 75.4% on DeepSWE v1.1, 59.4% on SWEAtlas CodeBase QnA, 88.8% on Terminal-Bench 2.1 (tied with [GPT Sol 5.6](/openai/gpt-sol/5-6)), and 98.5% / 98.1% on MRCR at 256K-512K and 512K-1M. Standard API pricing is $1.25/M input and $4.25/M output ($0.15/M cache reads), matching Cursor's listed rates. Measured price per task is $1.60 at max reasoning.
Anthropic's most advanced generally available model for coding, knowledge work, and long-running problem-solving. Same underlying weights as [Claude Mythos 5.1](/anthropic/mythos/5-1), with production safeguards. More than doubles [Claude Fable 5](/anthropic/fable/5) on Terminal-Bench-Science 0.1 (52.6% vs 24.7%) and AutomationBench (31.4% vs 17.1%), and leads [Claude Opus 5](/anthropic/opus/5) and [GPT Sol 5.6](/openai/gpt-sol/5-6) on the published agentic coding and knowledge-work suite. List API pricing stays $10/M input and $50/M output; cache reads drop 75% to $0.25/M, cutting typical token-billed workloads by about 25% and highly agentic ones by up to about 45%. Available as claude-fable-5-1 on the Claude API, Claude apps, Amazon Bedrock, Google Cloud, and Microsoft Azure. 1M-token context, 128k max output, knowledge cutoff June 2026.
Same underlying model as [Claude Fable 5.1](/anthropic/fable/5-1) with more permissive safeguards for vetted cybersecurity and life sciences work. Restricted to Anthropic's Cyber Verification Program and Life Sciences Verification Program (initially a set of US organizations). Scores 60.9% on Terminal-Bench 4.0 versus Fable 5.1's 55.8% with production safeguards. Also powers Claude Security. Same list API pricing as Fable 5.1: $10/M input, $50/M output, and $0.25/M cache reads. API ID claude-mythos-5-1.
Generally available Gemini Omni model for conversational video generation and editing. Extends scenes using up to 10 seconds of prior context in 10-second increments out to 40 seconds, interpolates first and last frames, drafts at 360p up to 60% faster and about one-third the 720p cost, and upscales to 1080p or 4K. API ID gemini-omni-1.1-flash in Google AI Studio and Gemini Enterprise Agent Platform; also in Google Flow for AI Plus, Pro, and Ultra. Preview endpoint gemini-omni-flash-preview retires 2026-09-30. Input $1.50/M; video output $17.50/M (about $0.10/s at 720p, $0.03/s at 360p, $0.15/s at 1080p, $0.30/s at 4K).
Open-weight preview of the Qwen4 architecture: a 125B MoE (6B active) plus 51B n-gram embeddings, hybrid Gated DeltaNet and Qwen Sparse Attention, and native image and video input. 262K native context (extensible to 1M). Qwen reports 62.5 on SWE-bench Pro, 58.7 on DeepSWE v1.1, and 81.0 on SWE-bench Multilingual, ahead of [Qwen3.8-27B](/alibaba/qwen/3-8-27b) on long-horizon coding and agent tasks.
Google's most precise speech-to-text model yet, succeeding Chirp 3 for live captioning, voice agents, and file transcription. File API gemini-3.5-transcribe adds speaker attribution and word-level timestamps; Live API gemini-3.5-transcribe-live streams with sub-second latency. Published results show 4.0% WER streaming and 2.6% non-streaming; FLEURS top-locale streaming WER is 5.50% versus 7.32% for Chirp 3. Supports 85+ languages, custom vocabulary, and filler-word cleanup. Public preview in Google AI Studio, Antigravity, and Gemini Enterprise. File pricing $2/M audio input and $12/M text output (about $0.005/min); live is $3.50/$21 (about $0.009/min).
First natively multimodal model in the GLM-5 line: 320B total parameters with 18B active, hybrid sparse plus linear attention, and MIT open weights. Outperforms [GLM-5.2](/zhipu/glm/5-2) on DeepSWE v1.1 (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2) at Flash API pricing ($0.15/M input, $0.50/M output) with a 1M-token context.
Dense 27B vision-language sibling of [Qwen3.8-Max](/alibaba/qwen/3-8-max) for local deployment. Apache 2.0 weights, 262K native context (extensible to 1M), image and video input, and 73.0 on Terminal-Bench 2.1 / 61.7 on SWE-bench Pro in Qwen's post-training evals.
Post-training upgrade on the same base as [GLM-5.2](/zhipu/glm/5-2). Z.ai reports a 50% gain on its in-house Code Bench, open-weight SOTA on Terminal-Bench 3.0 (28.3 vs 4.6) and DeepSWE v1.1 (66.9 vs 46.2), plus emergent cyber capability (CyberGym 84.5%, ExploitBench 54.4%). Weights shipped two weeks later under the GLM-5.3 License.
Google's most intelligent Flash workhorse yet for coding and agents, released three weeks after [Gemini 3.6 Flash](/google/gemini/3-6-flash). Strong gains on FrontierCode 1.1 Main (43.6%), DeepSWE v1.1 (65.3%), Arena.ai WebDev Arena (Elo 1588), GDP.pdf (34.0%), and AutomationBench (30.4%). Introductory API pricing at half the original 3.6 Flash cost ($0.75/M input, $3.75/M output through 2026-12-31). Powers Gemini Spark for Google AI Pro and Ultra subscribers.
Open-weight checkpoint of [Qwen3.8-Max](/alibaba/qwen/3-8-max): 2.4T total / 95B active sparse MoE, text-only, 262K native context (extensible toward 1M) under the Qwen3.8-Max License. Published on Hugging Face and ModelScope as Qwen3.8-2.4T-A95B.
SpaceXAI's successor to [Grok 4.5](/xai/grok/4-5), jointly trained with Cursor for long-running agents and visual work. Matches GPT-5.6 Sol on a broad independent intelligence benchmark (61 at high reasoning), scores 65.9% on DeepSWE v1.1 and 69.9% on CursorBench v3.2, and ships in Cursor, Grok Build, and the API as grok-4.6 at $2/$6 per million tokens (Fast is 2x).
Updated GitHub Copilot coding workhorse after [MAI-Code-1-Flash](/microsoft/mai-code/1-flash). Microsoft reports 25% greater token efficiency, 25% faster streaming, a 22% gain on Terminal-Bench 2.1 in Copilot CLI, and a 15% gain on .NET tasks, at one quarter the serving cost of the June model. Native screenshot and diagram understanding, in production in GitHub Copilot.
NVIDIA's open 31.6B-total, 3.6B-active hybrid MoE for high-volume agent execution. Successor to [Nemotron 3 Nano](/nvidia/nemotron/3-nano) in the small-model slot: OpenMDW-1.1 weights, 1M-token context, NVFP4 plus DFlash/DSpark speculative decoding, and 24.58 on Terminal-Bench 2.1 versus Nano's 8.29. NVIDIA positions it as the fast execution layer beside [Nemotron 3 Ultra](/nvidia/nemotron/3-ultra) for planning.
Meta's 30B open-weight multimodal agent model distilled from [Muse Spark](/meta/muse-spark/1-1) for local coding, tool use, and always-on device agents. Apache 2.0 weights, 131K context, text and image input, and a DFlash drafter for speculative decoding on a single consumer GPU. Meta reports 51.2 on SWE-bench Pro, 76.0 on SWE-bench Verified, and 51.7 on Terminal-Bench 2.1 at high reasoning versus [Gemma 4](/google/gemma/4) 31B and [Qwen3.6-27B](/alibaba/qwen/3-6-27b).
SpaceXAI's production image model for generation and editing, generally available as Quality Mode on grok.com/imagine and in Grok iOS and Android apps. Treats editing as a first-class capability: magic wand and segmentation for region changes, background removal with transparency, up to five reference images, and smart resize across aspect ratios. SpaceXAI reports it ranks second on Arena text-to-image and image-editing leaderboards as of 2026-08-07, behind OpenAI's GPT-Image-2, the predecessor to [GPT Image 2.5 Flare](/openai/gpt-image/2-5-flare). API ID grok-imagine-image-2.0 at $0.04 per output image.
Alibaba's 2.4T-parameter (95B active) multimodal flagship after [Qwen3.7-Max](/alibaba/qwen/3-7-max). 1M-token context, native text/image/video input, and stronger long-horizon coding and cowork: 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench, and 86.1 on OSWorld-Verified. Served as qwen3.8-max on QwenCloud at $2/$6 per million tokens.
Major Opus-tier upgrade that approaches [Claude Fable 5](/anthropic/fable/5) frontier intelligence at half the price. State-of-the-art on coding and knowledge work evaluations including Frontier-Bench and GDPval-AA, with strong gains in agentic thoroughness, visual outputs, and scientific research. New default on Claude Max and strongest model on Claude Pro. Supports up to a 1M-token context window. Same API pricing as [Claude Opus 4.8](/anthropic/opus/4-8) at $5/M input and $25/M output, plus Fast mode at 2.5x speed. Available as claude-opus-5 on the Claude API.
Public-preview Flash tier of [MAI Voice 2](/microsoft/mai-voice/2) for low-latency agents, assistants, and call-center speech. Microsoft reports it is 2x faster than Voice 2 and 32% cheaper, at $15 per 1M characters versus $22. Model-inference latency for 45 seconds of audio is 225ms versus 1s on Voice 2. Same 15 languages, 18 locales, gated instant voice cloning, and emotion control. Live on Microsoft Foundry, Azure Speech, Azure Voice Live, MAI Playground, and Dynamics 365 Contact Center.
Poolside's 118B-total, 8B-active open-weight coding MoE, scaled from the [Laguna XS 2.1](/poolside/laguna/xs-2-1) recipe to a 1M-token context. 70.2 on Terminal-Bench 2.1, 59.4 on SWE-bench Pro, and 40.4 on DeepSWE v1.1 in Poolside's harness, matching models many times its size. OpenMDW-1.1 weights, thinking and no-thinking modes, and OpenRouter pricing of $0.10/$0.20 per million tokens.
Moonshot AI's most capable Kimi model to date: 2.8 trillion parameters with Kimi Delta Attention (hybrid linear attention) and Attention Residuals, native visual understanding (text, image, and video input), and a flat 1M-token context window. Always-on thinking via reasoning_effort (currently max only). Aimed at long-horizon software engineering, knowledge work, and deep reasoning. API model ID kimi-k3 on the Moonshot OpenAI-compatible endpoint. Flat pay-as-you-go pricing with automatic context caching: $0.30/M cache-hit input, $3/M cache-miss input, and $15/M output (same list rates as Anthropic's [Claude Sonnet](/anthropic/sonnet/5) series, close to [GPT Terra 5.6](/openai/gpt-terra/5-6) at $2.5/$15). Launch reporting places it competitive with current frontier models and behind [Claude Fable 5](/anthropic/fable/5) and [GPT Sol 5.6](/openai/gpt-sol/5-6). Real cost still depends on reasoning-token efficiency versus leaner models like Sol. Available on kimi.com, Kimi Code, and the Kimi API; public weight downloads were not yet published at launch.
Meta's updated multimodal reasoning model for coding, tool use, computer interaction, and multi-agent workflows. [Muse Spark 1](/meta/muse-spark/1) gains a 1M-token context window and broader availability through the Meta Model API public preview and Meta AI's Thinking mode.
SpaceXAI's frontier model for coding, agentic tasks, and knowledge work, succeeding [Grok 4.1 Fast](/xai/grok/4-1-fast). Grok 4.5 runs at up to 80 tokens per second, uses roughly half as many tokens as comparable models on agentic tasks, and is available in Grok Build, Cursor, and the API.
OpenAI's premier full-duplex voice model for real-time conversations. It listens and speaks at the same time, handles interruptions without waiting for a turn to end, and delegates reasoning and tool use to a backend agent such as [GPT Terra 5.6](/openai/gpt-terra/5-6). First shipped in ChatGPT Voice (paid default; GPT-Live-1 mini is the free-tier sibling), then launched in the API as gpt-live-1. Voice sessions cost $0.05 per minute, billed per second; backend model and tool usage is billed separately. Speak reported almost 80% fewer interruptions versus previous turn-based systems.
Poolside's upgraded 33B-total, 3B-active local coding model improves multilingual and terminal-style agent work while retaining a 256K context window. It ships with open weights under OpenMDW 1.1 and paid API pricing of $0.10/M input and $0.20/M output tokens.
Anthropic's most agentic Sonnet yet: top-tier coding and tool use at Sonnet pricing with a 1M context window. Performance is close to [Claude Opus 4.8](/anthropic/opus/4-8) on agentic benchmarks, with a substantial upgrade over [Claude Sonnet 4.6](/anthropic/sonnet/4-6) on reasoning, tool use, coding, and knowledge work. Default model on Free and Pro plans and new default in Claude Code for Pro users. Introductory API pricing of $2/M input and $10/M output through August 31, 2026, then standard $3/$15; available as claude-sonnet-5 on the Claude Platform.
OpenAI's strongest model yet and flagship of the new GPT 5.6 tier naming: the generation number (5.6) advances on its own cadence while Sol, Terra, and Luna are durable capability tiers. State-of-the-art on Terminal-Bench 2.1 for agentic command-line coding; stronger GeneBench v1 genomics results than [GPT 5.5](/openai/gpt/5-5) with fewer tokens; on ExploitBench² competitive with [Claude Mythos Preview](/anthropic/mythos/preview) using roughly one-third of the output tokens. Introduces max reasoning effort and ultra mode (parallel subagents). Ships with OpenAI's most robust safety stack to date. Limited preview via API and Codex for trusted partners; broader ChatGPT and API availability planned in coming weeks. API pricing $5/M input and $30/M output, matching the prior [GPT 5.5](/openai/gpt/5-5) flagship. Positioned against [Claude Fable 5](/anthropic/fable/5) and [Claude Mythos 5](/anthropic/mythos/5) on frontier agentic and cyber workloads.
OpenAI's balanced GPT 5.6 tier for everyday work: performance competitive with [GPT 5.5](/openai/gpt/5-5) at half the API cost ($2.50/M input, $15/M output). Part of the durable Sol/Terra/Luna family where Terra sits between flagship [GPT Sol 5.6](/openai/gpt-sol/5-6) and affordable [GPT Luna 5.6](/openai/gpt-luna/5-6). Strong improvements on ExploitGym cyber benchmarks as reasoning effort increases. Analogous to Anthropic's balanced frontier lines such as [Claude Opus 4.8](/anthropic/opus/4-8) and [Claude Sonnet 4.6](/anthropic/sonnet/4-6) for cost-performance everyday agentic work, while Sol targets [Claude Fable 5](/anthropic/fable/5) / [Mythos 5](/anthropic/mythos/5) territory. Limited preview via API and Codex with general availability planned soon.
OpenAI's fast, affordable GPT 5.6 tier: strong capability at the lowest cost in the family ($1/M input, $6/M output). Completes the durable Sol/Terra/Luna lineup alongside [GPT Sol 5.6](/openai/gpt-sol/5-6) (flagship) and [GPT Terra 5.6](/openai/gpt-terra/5-6) (balanced). All three GPT 5.6 tiers show strong ExploitGym cyber improvements as reasoning scales. Positioned as OpenAI's volume tier, comparable in role to [Claude Haiku 4.5](/anthropic/haiku/4-5) for low-latency, cost-sensitive workloads, while [GPT 5.5](/openai/gpt/5-5) remains the prior-generation unified flagship on the legacy GPT product line. Limited preview via API and Codex.
SpaceXAI's video model, generally available on the Imagine API as grok-imagine-video-1.5. Sound effects, ambience, and dialogue are generated in the same pass as the picture. The June 16, 2026 release also rolled out Video 1.5 Fast on grok.com/imagine and the Grok iOS and Android apps, producing a 6-second 720p clip in about 25 seconds, down from 40+ seconds on the previous model. A July 31 update added text-to-video, up to seven image references, preset voice references, and native 1080p for text-to-video and image-to-video. Reference-to-video stays capped at 720p. Clips run 1 to 15 seconds. Listed rates are $0.08 per second at 480p, $0.14 at 720p, and $0.25 at 1080p, plus $0.01 per input image. The model page headline of $0.080 per second matches the 480p default. Available in us-east-1, us-west-2, and us-saltlake-2, with the Batch API and a 10 request per second cap.
Zhipu AI's open-weight flagship for long-horizon coding after [GLM-5](/zhipu/glm/5). Adds a usable 1M-token context via IndexShare sparse attention, MIT-licensed 753B MoE weights (40B active), and a jump to 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro. Rolled out first on the GLM Coding Plan, then the API and Hugging Face.
Experimental open Gemma model that generates text via diffusion instead of autoregressive decoding. 26B MoE (3.8B active) built on Gemma 4 and Gemini Diffusion research, released under Apache 2.0. Delivers up to 4x faster token generation on dedicated GPUs (1000+ tok/s on H100, 700+ on RTX 5090) by drafting 256-token blocks in parallel with bi-directional attention. Fits in 18GB VRAM when quantized. Best for speed-critical local workflows like in-line editing, code infilling, and rapid iteration; standard Gemma 4 remains recommended for maximum output quality.
Anthropic's first generally available Mythos-class model: state-of-the-art on software engineering, knowledge work, vision, and scientific research, with the largest lead over prior Claude models on long, complex tasks. Ships with safety classifiers that route some cybersecurity, biology, and distillation queries to Opus 4.8 (triggering in under 5% of sessions on average). Available on the Claude API as claude-fable-5, Claude apps, and major cloud platforms at $10/M input and $50/M output.
Same underlying model as [Claude Fable 5](/anthropic/fable/5) with cyber safeguards lifted for vetted partners. Strongest cybersecurity capabilities of any model in the world at launch, deployed through Project Glasswing as an upgrade to Mythos Preview. Restricted to Glasswing partners initially, with a broader trusted access program planned. API pricing at $10/M input and $50/M output.
Cohere's first dedicated agentic coding model is a 30B-total, 3B-active open MoE for repository changes, terminal work, and local deployment. It ships with a 256K context window, 64K maximum output, tool use, structured output, and Apache 2.0 weights.
NVIDIA's open 550B-total, 55B-active reasoning engine is the largest Nemotron 3 model and is designed for long-running agents, complex planning, multi-step tool use, coding, and enterprise workflows with up to 1M tokens of context.
Microsoft AI's flagship MoE reasoning model (~35B active, ~1T total parameters) for math, coding, and enterprise workloads. Competitive with Claude Opus 4.6 on SWE-Bench Pro at a smaller inference footprint, preferred to Sonnet 4.6 in blind Surge human evals, with built-in safety guardrails and copyright protection. Available in Microsoft Foundry private preview.
5B-parameter inference-efficient agentic coding model custom-trained for GitHub Copilot and VS Code. Plans and reasons through multi-step coding tasks, supports broad language ecosystems, and is positioned as comparable to Claude Haiku at lower cost. Rolling out in GitHub Copilot in VS Code.
Text-to-image and image-editing model for photorealistic, design-ready output with fine-grained edit control, reliable text rendering, and branding or product workflows. Includes an ultra-efficient Flash variant; Microsoft reports Arena scores surpassing Nano Banana Pro. Available via MAI Playground and Microsoft Foundry.
Speech-to-text model with 4.9% average WER on FLEURS across 43 languages (automatic detection), contextual biasing for domain terminology, and ~5.7x lower latency than cited competitors. Outperforms Scribe v2, Whisper-large-v3, GPT-4o-Transcribe, and Gemini 3.1 Flash on many language benchmarks. Priced at $0.36 per hour via Azure Speech / Foundry.
Multilingual text-to-speech with 15 languages, instant voice matching from short reference clips, expressive emotion control, and stable long-form output for audiobooks, podcasts, and lectures. Built-in guardrails require authorized, consented voices. Priced at $22 per 1M characters via MAI Playground and Azure Speech.
12B-parameter Mixture-of-Experts focal model (2.5B active per token) for routing, RAG, sub-agents, and private deployments. Ships open from day one with base, instruct, and thinking checkpoints under Apache 2.0.
MiniMax's open-weight multimodal frontier model combines coding, agentic computer use, image and video input, and long-context work. Its MiniMax Sparse Attention architecture supports up to one million tokens while improving prefill and decoding efficiency over the M2 generation.
Upgrade to Anthropic's Opus class with stronger performance across coding, agentic tasks, and professional work, plus improved consistency for long-running tasks. Adds effort control in claude.ai and Cowork, dynamic workflows in Claude Code (research preview) for large parallel subagent runs, and fast mode at 2.5x speed with pricing three times lower than on prior Opus models. Same standard API pricing as Opus 4.7 ($5/M input, $25/M output). Available via Claude API as claude-opus-4-8.
Cohere's 218B-total, 25B-active open MoE unifies reasoning, vision, multilingual work, and tool use in one enterprise model. It supports 48 languages, 128K input, 64K output, and Apache 2.0 weights that can run on a single B200 or two H100 GPUs.
Gemini 3.5 Flash is Google's agentic and coding-focused Flash model: frontier-class scores on Terminal-Bench 2.1 (76.2%), GDPval-AA, and MCP Atlas at roughly 4x the output tokens per second of other frontier models.
Cursor's updated frontier coding model for Composer, trained with scaled RL, harder synthetic tasks, and targeted textual feedback. Strong gains over Composer 2 on Terminal-Bench 2.0 (69.3%), SWE-bench Multilingual (79.8%), and CursorBench v3.1 harder tasks (63.2%). Same standard pricing as Composer 2 ($0.50/M input, $2.50/M output); a faster default variant costs more ($3.00/M input, $15.00/M output). Double usage for the first week after launch.
NVIDIA's open 30B-total, 3B-active omni model combines reasoning over text, images, video, and audio for document intelligence, computer use, and multimodal agents. NVIDIA released the model with open weights, training data, and recipes, plus a 256K context window and 64K maximum output.
Poolside's 225B-total, 23B-active flagship MoE for long-horizon coding agents launched alongside the Laguna runtime and agent experiences. It targets multi-file debugging, repository exploration, tool use, and validated software changes with a 256K served context window.
Poolside's compact 33B-total, 3B-active coding MoE brought the Laguna family to local machines with Apache 2.0 open weights. It supports reasoning, tool use, a 256K served context window, and long-horizon software engineering workflows.
The flagship variant of the [DeepSeek V4](/deepseek/deepseek/v4) family, with 1M-token context, frontier reasoning, long-context coding, and competitive scores on agentic engineering benchmarks.
Cost-efficient fast tier of [DeepSeek V4](/deepseek/deepseek/v4) with a 1M-token context window, now officially released with much stronger agentic coding. The July 2026 upgrade reaches 82.7 on Terminal-Bench 2.1 while keeping the same 284B-parameter MoE footprint.
DeepSeek's open-source V4 Preview family introduces cost-efficient 1M-token context for long-horizon coding and agentic work. It succeeds [DeepSeek V3.2](/deepseek/deepseek/v3-2) and ships as the fast [V4-Flash](/deepseek/deepseek/v4-flash) and flagship [V4-Pro](/deepseek/deepseek/v4-pro) variants.
OpenAI's smartest general-purpose frontier model for agentic coding, computer use, and knowledge work. State-of-the-art on Terminal-Bench 2.0 (82.7%), SWE-Bench Pro (58.6%), Expert-SWE (73.1%), GDPval (84.9%), OSWorld-Verified (78.7%), and BrowseComp (84.4%). Matches GPT-5.4 per-token latency while using fewer tokens on Codex tasks. Rolling out in ChatGPT and Codex; API pricing at $5/M input and $30/M output with 1M context.
Open-weight coding and agent model from Moonshot AI. Reported SOTA among open models on Humanity's Last Exam with tools (54.0%), SWE-Bench Pro (58.6%), and strong scores on SWE-bench Multilingual (76.7%), BrowseComp (83.2%), Toolathlon (50.0%), CharXiv with Python (86.7%), and MathVision with Python (93.2%). Positioned for GPT-5.4-class coding at much lower cost than closed frontier APIs, with long-horizon runs (4,000+ tool calls, 12+ hours), large agent swarms (300 parallel sub-agents, 4,000 steps per run), and multimodal front-end capabilities. Available on kimi.com (chat and agent mode); Kimi Code at kimi.com/code targets production workflows.
Meta's first Muse family model, a natively multimodal reasoning system for visual chain-of-thought, tool use, coding, and multi-agent orchestration. It complements Meta's open [Llama 4 Maverick](/meta/llama/4-maverick) line with an agent-focused model available through Meta AI and an API preview.
First Mythos-class frontier model, announced through Project Glasswing for defensive cybersecurity with AWS, Google, Microsoft, and other partners. Not generally available: restricted to vetted critical-infrastructure and security organizations. Demonstrated autonomous vulnerability discovery surpassing skilled humans; priced at $25/M input and $125/M output for partners.
Cursor's frontier coding model for Composer, trained with reinforcement learning (including compaction-in-the-loop) for long-horizon tasks. Announced with strong gains on CursorBench, Terminal-Bench 2.0, and SWE-bench Multilingual. Standard API-style pricing plus a faster default variant; available in Cursor and the early Glass alpha.
OpenAI's most capable frontier model with 1M token context (Codex/API), native computer-use, and top benchmarks: SWE-Bench Pro 57.7%, OSWorld-Verified 75% (beats human 72.4%), BrowseComp 82.7%, GDPval 83%, Toolathlon 54.6%. More token-efficient and faster reasoning than 5.2.
Google's latest Gemini model designed for complex tasks where a simple answer isn't enough. Builds on the Gemini 3 family with enhanced reasoning and deeper analysis capabilities for sophisticated computational challenges.
Full upgrade across coding, computer use, long-context reasoning, agent planning, and knowledge work. Features 1M token context window (beta) and significant improvements in consistency and instruction following. Users prefer it to Sonnet 4.5 roughly 70% of the time, and over Opus 4.5 59% of the time — approaching Opus-level intelligence at $3/$15 per million tokens.
OpenAI's fastest coding model yet. Built on the Codex line, Codex Spark prioritizes speed and low-latency responses while retaining strong coding performance - ideal for real-time pair-programming and interactive development workflows.
Google's specialized reasoning mode for science, research, and engineering. Uses iterative rounds of reasoning to explore multiple hypotheses simultaneously. Achieves 48.4% on Humanity's Last Exam (without tools) and 84.6% on ARC-AGI-2. Demonstrates gold medal-level results on the 2025 International Physics and Chemistry Olympiads.
Zhipu AI's open-weight 744B parameter MoE model (40B active per token) trained on 28.5T tokens. First open model to cross 50 on a broad independent intelligence benchmark, outperforming Gemini 3.0 Pro and GPT-5.2 across multiple evaluations. Released under MIT License.
Thinking agentic coding model from Cursor with 20x scaled RL on the Composer 1 base, self-summarization for long tasks, and adaptive depth (minimal thinking on easy prompts, extended reasoning on hard ones).
Major update with 1M token context window (beta), improved coding and debugging capabilities, enhanced reasoning with adaptive thinking, and new agent teams feature in Claude Code.
OpenAI's latest coding-focused model with best-in-class coding performance (57% SWE-Bench Pro, 76% TerminalBench 2.0, 64% OSWorld). Features customizable personality modes ('pragmatic' or 'friendly'), mid-task steerability with live updates, and the first model to achieve 'high' rating for cybersecurity on OpenAI's preparedness framework. Significantly faster and more efficient—uses less than half the tokens of 5.2-Codex for equivalent tasks and runs >25% faster per token.
NVIDIA's open 30B-total, 3B-active hybrid MoE brought efficient reasoning, coding, retrieval, and agent workflows to the Nemotron 3 family. It ships with disclosed training resources and supports a 1M-token context window.
Anthropic's coding- and agent-focused Sonnet with checkpoints in Claude Code, computer-use gains, code execution in Claude apps, and the same $3/$15 pricing as Sonnet 4.
OpenAI's unified GPT-5 system with adaptive reasoning, state-of-the-art performance across coding, math, writing, and health, available across ChatGPT tiers and the API.
OpenAI API-only GPT-4.1 family (4.1, 4.1 mini, 4.1 nano) with 1M-token context, stronger coding and instruction following than GPT-4o, and June 2024 knowledge cutoff.
Cohere's 111B open-weight enterprise model for agentic workflows, retrieval, multilingual work, and tool use. Command A launched with a 256K context window and was designed to run on two GPUs while competing with substantially larger models.
OpenAI's largest GPT chat model at launch, scaled for broader knowledge, improved alignment, and more natural conversation with fewer hallucinations than GPT-4o, without o-series chain-of-thought reasoning.
Anthropic's first hybrid reasoning Sonnet with visible extended thinking, strong coding and agentic gains, and fine-grained thinking budget control in the API.
Experimental Gemini 2.0 Pro focused on agentic capabilities, native multimodality, and faster responses for developers on Vertex AI and Google AI Studio.
Meta's 70B instruct Llama with post-training refinements that match Llama 3.1 405B quality on many benchmarks at far lower serving cost, with 128K context.
JetBrains' first focal model: a 4B-parameter LLM trained from scratch for multilingual cloud code completion in JetBrains IDEs, with deep IDE integration for project-aware suggestions.
Qwen2.5 expanded the open family with stronger coding, mathematics, structured output, instruction following, and multilingual support across 29 languages.
Meta's first frontier-scale open-weights Llama with 405B parameters, 128K context, multilingual support, and tool use, rivaling closed models on many benchmarks.
First Claude 3.5 model: outperforms Claude 3 Opus on many benchmarks at Sonnet-tier speed and cost, with Artifacts for editable code and documents in the Claude app.
OpenAI's flagship omni model reasoning across audio, vision, and text in real time, rolled out in ChatGPT (including free tier) with improved latency versus GPT-4.
Qwen1.5 open-sourced base and chat models across a broad size range with 32K context, stronger multilingual alignment, and standard Transformers support.
Google made Gemini 1.0 Pro available to developers through Google AI Studio and Vertex AI as the balanced tier of its first natively multimodal family.
OpenAI's first large multimodal GPT-4 model, with strong performance on professional and academic benchmarks and broad availability in ChatGPT Plus and via API waitlist.