Know if it has beenreleased.Instantly.

Live status and latest versions for the AI models you care about. No more hunting through announcement feeds or refreshing Twitter.

Latest actionable insights

What changed

The practical decisions hidden inside the latest model releases.

Briefing updated Nov 2

Long-horizon defenseOct 28

$7.76

price per task

Hold Google Gemini 4 Carbon for Fairwind cyber reviews

Google Gemini 4 Carbon lists at an introductory $2/$10 per million tokens, then $4/$20, with cached input 95% below the input rate. A measured price per task at high effort is $7.76 at the standard $4/$20 rate. On Google's comparison table, SEC-bench Pro is 88.2%, ExploitGym is 54.8%, and CWE-bench is 71.0%. Terminal-Bench 4.0 is 65.2%. Anthropic Claude Opus 5.5 leads OSWorld 2.1 on that table at 87.2%, against 81.5% for Gemini 4 Carbon. OpenAI GPT Astra 6 leads FrontierSWE 2 at 65.5%, against 61.2%. Access starts with cyber defenders in the Fairwind Program, and Google has not published an API model ID.

Your move: Use Google Gemini 4 Carbon for a bounded defensive review when you have Fairwind access and $7.76 per task at the standard rate is acceptable. Keep Anthropic Claude Opus 5.5 when this table's OSWorld 2.1 score of 87.2% is the bar, and OpenAI GPT Astra 6 when FrontierSWE 2 needs 65.5%.

Google Gemini 4 CarbonDetails

Speech transcriptionOct 28

4.71%

Open ASR mean WER

Transcribe long recordings on NVIDIA Nemotron Transcribe 3.5

NVIDIA Nemotron Transcribe 3.5 is an open-weight speech model, 34B total and about 4.3B active, under OpenMDW-1.1. The model card documents a 65,536-token context and recordings up to one hour. Mean word error rate is 4.71% on the nine public English Open ASR sets and 4.62% on the multilingual Open ASR leaderboard. Speaker-attributed cpWER on full sessions is 15.80%. No first-party token price is published. Mistral Voxtral Small Transcribe 3 remains the hosted file model at $0.004 per minute.

Your move: Use NVIDIA Nemotron Transcribe 3.5 when recorded audio has to stay on your own machines and the transcript needs speaker labels. Keep Mistral Voxtral Small Transcribe 3 when you want a hosted file API at $0.004 per minute.

NVIDIA Nemotron Transcribe 3.5Details

Computer useOct 26

$0.96

price per task

Send high-volume computer use to OpenAI GPT-6.1 Terra

OpenAI GPT-6.1 Terra lists at $1/$5 per million tokens, with cache reads at $0.05. A measured price per task at max effort is $0.96, and medium, the API default, is $0.17. Agents' Last Exam v1 at max is 60.0%, against 63.2% for Anthropic Claude Opus 5.5. Terminal-Bench 4.0 at max is 50.3%, more than double 5.6's published 21.5%. GPT-6.1 Sol scores 55.9% on that benchmark.

Your move: Use OpenAI GPT-6.1 Terra for high-volume visual and computer-use work when $0.96 per task at max is enough. Keep GPT-6.1 Sol when Terminal-Bench 4.0 needs its 55.9% score, and Anthropic Claude Opus 5.5 when Agents' Last Exam v1 needs 63.2%.

OpenAI GPT Terra 6.1Details

Frontier codingOct 23

$2.11

price per task

Price frontier coding with Moonshot AI Kimi K3.1 Pro

Moonshot AI Kimi K3.1 Pro lists at $2/$10 per million tokens, with cache reads at $0.20. A measured price per task is $2.11. Terminal-Bench 4.0 is 65.7% with the Claude Code harness at max effort, ahead of OpenAI GPT-6 Astra at 58.2% and behind Anthropic Claude Opus 5.5 at 66.4%. SWE-Marathon v1.1 is 51.7% with the same harness. Highspeed is the same checkpoint at 2-3x speed for $4/$20.

Your move: Send long coding-agent jobs to Moonshot AI Kimi K3.1 Pro when $2.11 per task is enough. Keep Anthropic Claude Opus 5.5 when its 66.4% Terminal-Bench 4.0 score is the bar, and MiniMax M3.1 Flash when the budget is about $0.22 per task.

Moonshot AI Kimi K3.1 ProDetails

Rumored Releases

Releases everyone is waiting for

Llama 4 Behemoth

byMeta

TBD

Meta previewed Llama 4 Behemoth in the April 2025 Llama 4 post as a 288B-active, about 2T-total teacher model. That post published STEM numbers while stating the model was still training and not being released. It is not a downloadable or API preview.

View details

Recent Releases

Latest versions that shipped

Mage VL Cowork

byMicrosoft AI

Oct 30

Microsoft's open-weight computer-use models that act through the screen and a persistent terminal in one checkpoint.

Release summary: Microsoft's open-weight 4B computer-use model. It acts through screenshots and a persistent terminal in one checkpoint, with no external orchestrator. Apache 2.0 weights at microsoft/Mage-VL-Cowork. Image and text in, text out. The model card documents a 32,768-token inference window. The architecture supports up to 262,144 tokens, and that longer ceiling is not the published inference window, so this catalog uses 32,768 for both context and output. On OSWorld-Verified, with the MAGEHybridAgent harness, a 50-step budget, and 361 tasks excluding Google Drive, the two-run average is 66.0% (65.33% and 66.75%). Category scores on that table are 87.5% for OS, 64.9% for Office, 65.0% for Daily, 66.3% for Professional, and 62.7% for multi-application workflows. [Muse Glimmer 30B](/meta/muse-glimmer/30b) is 65.9% on the same table. The released checkpoint uses 18.77 model calls and 5,499 output tokens per task on that evaluation. No hosted token price and no measured price per task.

View details

Gemini 4 Carbon

byGoogle

Oct 28

Google's multimodal AI model series built by DeepMind.

Release summary: Google's next frontier Gemini model after [Gemini 4 Argon](/google/gemini/4-argon), for cybersecurity defense and long-horizon engineering. Introductory list price is $2 per million input tokens and $10 per million output tokens, with cached input 95% below the input rate. After the introductory period the list price is $4/$20. The end date is not published. A measured price per task at high effort is $7.76 at the standard $4/$20 rate. The announcement does not list a context or output cap. This catalog uses a 1M-token combined window for both. Text, image, video, and speech in, text out. Google has not published an API model ID. External access starts with a subset of cyber defenders through the Fairwind Program. On Google's comparison table, SEC-bench Pro is 88.2%, ExploitGym is 54.8%, CWE-bench is 71.0%, and Terminal-Bench 4.0 is 65.2%. [Claude Opus 5.5](/anthropic/opus/5-5) leads OSWorld 2.1 on that table at 87.2%, against 81.5% here. [GPT Astra 6](/openai/gpt-astra/6) leads FrontierSWE 2 at 65.5%, against 61.2%. WebDev Arena is 1768 Elo. Gray Swan indirect prompt injection attack success at 15 attempts is 0.0%.

View details

Nemotron Transcribe 3.5

byNVIDIA

Oct 28

NVIDIA's open-weight speech models for promptable transcription, translation, and speaker attribution on recorded audio.

Release summary: NVIDIA's open-weight speech model for recorded audio. Checkpoint NVIDIA-Nemotron-3.5-Transcribe-30B-A4B: 34B total parameters and about 4.3B active, including the audio encoder, built on the Nemotron 3.5 Lightning 30B hybrid language model. OpenMDW-1.1 weights. Text and audio in, text out. The model card documents a 65,536-token context shared by the prompt and the output, and recordings up to one hour. It transcribes 40 languages, translates 31 of them to or from English, and attributes up to eight speakers in 13 languages. Mean word error rate is 4.71% on the nine public English Open ASR sets, 4.62% on the multilingual Open ASR leaderboard, and 10.98% on the long-form Open ASR sets. Speaker-attributed mean cpWER on full AMI, CALLHOME, Mixer 6, and NOTSOFAR-1 sessions is 15.80%. Preference-ASR word error rate is 7.56% with the default style and 6.16% when the style is instructed. FP8, NVFP4, and GGUF checkpoints ship beside BF16. DFlash2 is a draft model for speculative decoding, not a separate release. The NVFP4 checkpoint with DFlash2, served by the Nemotron Transcribe NIM, transcribes up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell GPU. No first-party token price and no measured price per task. Hosted file model in this catalog: [Voxtral Small Transcribe 3](/mistral/voxtral/small-transcribe-3).

View details

GPT Terra 6.1

byOpenAI

Oct 26

OpenAI's balanced tier for everyday knowledge work and agentic tasks, positioned between Sol and Luna on intelligence, speed, and cost.

Release summary: GPT-6.1 generation of OpenAI's Terra tier for high-volume visual reasoning, computer use, and writing. Standard API rates are $1/M input, $0.05/M cached input, $1.25/M cache writes at 30-minute retention, and $5/M output. OpenAI describes those rates as 50% lower than [GPT Terra 5.6](/openai/gpt-terra/5-6) and [GPT Sol 6.1](/openai/gpt-sol/6-1). A measured price per task at max effort is $0.96; medium, the API default, is $0.17. Agents' Last Exam v1 at max is 60.0%, against 63.2% for [Claude Opus 5.5](/anthropic/opus/5-5). At medium effort it is 15 points above [GPT Terra 5.6](/openai/gpt-terra/5-6) at about one-third the charted cost. Terminal-Bench 4.0 at max is 50.3%, more than double that version's published 21.5%, and below [GPT Sol 6.1](/openai/gpt-sol/6-1) at 55.9%. Terminal-Bench-Science 0.1 at max is 58.3%, above [GPT Sol 6.1](/openai/gpt-sol/6-1) at 57.0% and within five points of [Claude Opus 5.5](/anthropic/opus/5-5) with fallbacks. Available as gpt-6.1-terra in the API, and in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu. Not in Chat. 1.05M context, up to 922k input, 128k max output, knowledge cutoff June 1, 2026. Text, image, and PDF in, text out. Reasoning effort is none, low, medium (the default), high, xhigh, or max. Fast mode is 2x Standard. Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the full request. Tool calling uses the Responses API. Uses the same safeguards stack as [GPT Astra 6](/openai/gpt-astra/6) and [GPT Sol 6.1](/openai/gpt-sol/6-1).

View details

Kimi K3.1 Pro

byMoonshot AI

Oct 23

Moonshot AI's open-weight model series for coding, agents, and long-horizon tasks.

Release summary: Post-trained [Kimi K3](/moonshot/kimi/k3) from Moonshot AI: 2.8 trillion total parameters and 104 billion active, with text, image, and video input. Context is 1,048,576 tokens. The API default output cap is 131,072 tokens and can be raised to the full context. API id kimi-k3.1-pro. Reasoning is always on, with reasoning_effort low, high, or max, and max is the default. List price is $2 per million input tokens, $0.20 cache reads, and $10 output. Cache writes are $2 for a 5-minute TTL and $4 for 1 hour. Highspeed (kimi-k3.1-pro-highspeed) is the same checkpoint at 2-3x speed for $4/$20, with the same $0.20 cache read. On the announcement chart at max effort, Terminal-Bench 4.0 is 65.7% with Claude Code, ahead of [GPT-6 Astra](/openai/gpt-astra/6) at 58.2% and behind [Claude Opus 5.5](/anthropic/opus/5-5) at 66.4%. SWE-Marathon v1.1 is 51.7% with the same harness. OfficeQA Pro v2 is 64.4. Measured price per task is $2.11. Weights are scheduled by November 4, 2026. Available on kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

View details

Nemotron 3.5 Bolt E4B

byNVIDIA

Oct 23

NVIDIA's open model family for efficient multimodal reasoning and enterprise agent systems.

Release summary: NVIDIA's open conversational model for local voice assistants, characters, and embedded agents. Post-trained from Gemma 4 E4B. 8 billion total parameters, 4 billion effective, text in and text out, and a 65,536-token context for both the prompt and the output. OpenMDW-1.1 weights. The Q4_K_M checkpoint is the 4GB release. On the BF16 card, MultiChallenge is 53.10, IFBench is 63.79, and BFCL-v3 is 68.51. The Q4_K_M checkpoint scores 55.60, 64.00, and 67.31 on those three. NVIDIA measures about 3.1GB to 4.3GB of memory and 52 to 133 tokens per second on RTX hardware at the short contexts in the post. A companion DFlash drafter is for speculative decoding, not a separate model. No hosted token price and no measured price per task. Smaller same-week sibling of [Nemotron 3.5 Bolt](/nvidia/nemotron/3-5-bolt).

View details

See all models

Know if it has beenreleased.Instantly.

Release cadence

Latest actionable insights

Hold Google Gemini 4 Carbon for Fairwind cyber reviews

Transcribe long recordings on NVIDIA Nemotron Transcribe 3.5

Send high-volume computer use to OpenAI GPT-6.1 Terra

Price frontier coding with Moonshot AI Kimi K3.1 Pro

Rumored Releases

Llama 4 Behemoth

Recent Releases

Mage VL Cowork

Gemini 4 Carbon

Nemotron Transcribe 3.5

GPT Terra 6.1

Kimi K3.1 Pro

Nemotron 3.5 Bolt E4B