agi.jetzt
DEEN

As of 15 Oct 2026, 11:38 UTC

New language models
at a glance.

language models in the registry
103
other endpoints: speech, image, embeddings, moderation
66
providers with language models · a frontier statement for 4
10
models in the pilot Prüfstand
6
Release map · language models · last 260 daysFrontier Set, as the provider positions ita preview, per the providertested in the Prüfstand
FebMarAprMayJunJulAugSeptOctOpenAIOpenAIGPT-5.4 · 5 Mar 2026 · not a Frontier modelGPT-5.4 · 05/03GPT-5.4 Pro · 5 Mar 2026 · not a Frontier modelGPT-5.4 Pro · 05/03GPT-5.4 mini · 17 Mar 2026 · not a Frontier modelGPT-5.4 mini · 17/03GPT-5.4 nano · 17 Mar 2026 · not a Frontier modelGPT-5.4 nano · 17/03GPT-5.5 · 23 Apr 2026 · not a Frontier modelGPT-5.5 · 23/04GPT-5.5 Pro · 24 Apr 2026 · not a Frontier modelGPT-5.5 Pro · 24/04GPT-5.6 Luna · 9 Jul 2026 · not a Frontier modelGPT-5.6 Luna · 09/07GPT-5.6 Sol · 9 Jul 2026 · not a Frontier modelGPT-5.6 Sol · 09/07GPT-5.6 Terra · 9 Jul 2026 · not a Frontier modelGPT-5.6 Terra · 09/07GPT-6 Astra · 3 Sept 2026 · FrontierGPT-6 Astra · 03/09GPT-6 Luna · 22 Sept 2026 · FrontierGPT-6 Luna · 22/09GPT-6 Sol · 22 Sept 2026 · not a Frontier modelGPT-6 Sol · 22/09GPT-6.1 Sol · 29 Sept 2026 · FrontierGPT-6.1 Sol · 29/09AnthropicAnthropicClaude Opus 4.6 · 5 Feb 2026 · not a Frontier modelClaude Opus 4.6 · 05/02Claude Sonnet 4.6 · 17 Feb 2026 · not a Frontier modelClaude Sonnet 4.6 · 17/02Claude Opus 4.7 · 16 Apr 2026 · not a Frontier modelClaude Opus 4.7Claude Opus 4.8 · 28 May 2026 · not a Frontier modelClaude Opus 4.8 · 28/05Claude Fable 5 · 9 Jun 2026 · not a Frontier modelClaude Fable 5 · 09/06Claude Mythos 5 · 9 Jun 2026 · not a Frontier modelClaude Mythos 5 · 09/06Claude Sonnet 5 · 30 Jun 2026 · not a Frontier modelClaude Sonnet 5 · 30/06Claude Opus 5 · 24 Jul 2026 · not a Frontier modelClaude Opus 5 · 24/07Claude Fable 5.1 · 1 Sept 2026 · FrontierClaude Fable 5.1 · 01/09Claude Mythos 5.1 · 1 Sept 2026 · not a Frontier modelClaude Mythos 5.1 · 01/09Claude Opus 5.5 · 22 Sept 2026 · FrontierClaude Opus 5.5 · 22/09Claude Sonnet 5.5 · 28 Sept 2026 · FrontierClaude Sonnet 5.5 · 28/09Claude Haiku 5.5 · 7 Oct 2026 · not a Frontier modelClaude Haiku 5.5 · 07/10GoogleGoogleGemini 3.1 Flash-Lite · 7 May 2026 · not a Frontier modelGemini 3.1 Flash-Lite · 07/05Gemini 3.5 Flash · 19 May 2026 · not a Frontier modelGemini 3.5 Flash · 19/05Gemini 3.5 Flash-Lite · 21 Jul 2026 · FrontierGemini 3.5 Flash-Lite · 21/07Gemini 3.6 Flash · 21 Jul 2026 · not a Frontier modelGemini 3.6 Flash · 21/07Gemini 3.7 Flash · 13 Aug 2026 · not a Frontier modelGemini 3.7 Flash · 13/08Gemini 3.8 Flash · 2 Sept 2026 · FrontierGemini 3.8 Flash · 02/09Gemini 4 Argon · 13 Oct 2026 · FrontierGemini 4 ArgonSpaceXAISpaceXAIGrok 4.5 · 16 Jul 2026 · not a Frontier modelGrok 4.5Grok 4.6 · 12 Aug 2026 · not a Frontier modelGrok 4.6Grok 4.7 · 21 Sept 2026 · FrontierGrok 4.7DeepSeekDeepSeekDeepSeek-V4-Pro · 24 Apr 2026 · not a Frontier modelDeepSeek-V4-ProDeepSeek-V4.1-Flash · 10 Sept 2026 · not a Frontier modelDeepSeek-V4.1-FlashMetaMetaMuse Spark · 8 Apr 2026 · not a Frontier modelMuse SparkMistral AIMistral AIMistral Small 4 · 16 Mar 2026 · not a Frontier modelMistral Small 4Mistral Medium 3.5 · 22 May 2026 · not a Frontier modelMistral Medium 3.5QwenQwenQwen3.8-Max · 2 Aug 2026 · not a Frontier modelQwen3.8-Max · 02/08Qwen3.8-27B · 14 Aug 2026 · not a Frontier modelQwen3.8-27B · 14/0815/10

← swipe · Dated releases of language models, from the registry: 43 in this window. 24 language models have no release date and cannot be placed; older and retired ones are in the registry further down. Highlighted is the Frontier Set, not the newest date.

Frontier now

The models each provider itself positions as its top tier — not the newest date. No ranking across providers. None of them has been tested in the Prüfstand yet.

  1. OpenAI

    • Highest capabilityGPT-6 Astra“our flagship model for complex reasoning and coding”Source: developers.openai.com · checked on 4 Oct 2026Release 3 Sept 2026 · Public benchmarks: 17 · no Prüfstand run
    • Work & codingGPT-6.1 Sol“Near-Astra performance for complex work at a lower cost”Source: developers.openai.com · checked on 4 Oct 2026Release 29 Sept 2026 · Public benchmarks: 8 · no Prüfstand run
    • Fast & efficientGPT-6 Luna“Our most efficient model for focused, high-volume tasks”Source: developers.openai.com · checked on 4 Oct 2026Release 22 Sept 2026 · Public benchmarks: 8 · no Prüfstand run

  2. Anthropic

  3. Google

    • Highest capabilityRestricted accessGemini 4 Argon“our new frontier model”Source: blog.google · checked on 10 Oct 2026Release 13 Oct 2026 · Public benchmarks: 3 · no Prüfstand run
    • Work & codingGemini 3.8 Flash“Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents”Source: ai.google.dev · checked on 4 Oct 2026Release 2 Sept 2026 · Public benchmarks: 15 · no Prüfstand run
    • Fast & efficientGemini 3.5 Flash-Lite“Our fastest, most cost-effective 3.5 model for high-throughput execution”Source: ai.google.dev · checked on 4 Oct 2026Release 21 Jul 2026 · Public benchmarks: 5 · no Prüfstand run

  4. SpaceXAI

    • Highest capabilityGrok 4.7“It is the most capable model we've built”Source: docs.x.ai · checked on 4 Oct 2026Release 21 Sept 2026 · Public benchmarks: 13 · no Prüfstand run

agi.jetzt makes no frontier statement about DeepSeek, Meta, Microsoft, Mistral AI, Moonshot AI, Qwen; their models are in the registry.

In the Prüfstand

6 models have public runs. Their runtimes differ and are shown as provenance.

  1. Anthropicmodel + harnessClaude Fable 56Prüfstand runs of 28Engineclaude-code-cliBenchmarks11 benchmarks recordedclaude-fable-5
  2. OpenAImodel + harnessGPT-5.6 Luna6Prüfstand runs of 28Enginecodex-cliBenchmarks5 benchmarks recordedgpt-5-6-luna
  3. Qwenmodel onlyQwen3.8-27B5Prüfstand runs of 28EngineollamaBenchmarksnone recordedqwen3-8-27b
  4. Anthropicmodel + harnessClaude Haiku 4.54Prüfstand runs of 28Engineclaude-code-cliBenchmarksnone recordedclaude-haiku-4-5
  5. OpenAImodel + harnessGPT-5.6 Sol4Prüfstand runs of 28Enginecodex-cliBenchmarks17 benchmarks recordedgpt-5-6-sol
  6. Microsoftmodel onlyPhi-4-mini-instruct3Prüfstand runs of 28EngineollamaBenchmarksnone recordedphi-4-mini-instruct

The model landscape

Not every relevant AI model is a generalist. The frontier models above are generalists. Other model classes have tasks of their own — and some of them work quite differently.

  • Generalists

    Broad models for reasoning, knowledge, coding, tool use and multimodal input. This is the class of the frontier models.

    Examples
    • GPT-6 AstraOpenAI
    • Claude Fable 5.1Anthropic
    • Gemini 4 ArgonGoogle
    • Grok 4.7SpaceXAI
    To the frontier models →

  • Image & Design

    Models that generate, edit and lay out images — increasingly with layers, layout and type under control.

    Examples
    • GPT-Image-2.5OpenAI
    • Nano Banana 2.1Google
    • FLUX 3 ImageBlack Forest Labs
    • Ideogram 4.5Ideogram
    Image & Design →

  • Video

    Models that generate moving images, often with sound. Their task is coherence over time, not one good frame.

    Examples
    • Gemini Omni FlashGoogle
    • Seedance 2.5ByteDance
    • Kling 4.0Kuaishou
    • FLUX 3 VideoBlack Forest Labs
    Video →

  • Speech & Audio

    Models that understand speech, speak, hold a conversation in real time or generate music.

    Examples
    • GPT-Live 1OpenAI
    • Gemini 3.8 LiveGoogle
    • Eleven v4ElevenLabs
    • MAI-Transcribe-2Microsoft
    Speech & Audio →

  • System One & Decision Models

    Models that return a choice, a score or a probability instead of text — for software, not for readers. TypeSafe AI calls its own "System One Models".

    1. State
    2. Decision
    3. Probability
    4. Action
    Examples
    • JevTypeSafe AI
    • Microsoft-Decision-1Microsoft
    • H2O-Lightning-4BH2O.ai
    • Quyet-1.0-LargeChinh Nguyen
    Decision Models →

These classes are not part of the frontier models and are not counted in AGI Progress. A selection with reasons, not a ranking.

The whole registry

Language models and the other endpoints, as the providers’ catalogs list them: search, filter and sort. 131 of 169 registry entries checked against a source. “Not verified” means on record without a source check — it does not mean unavailable.

The 14 most recently released of 103 language models. Search, filter or show everything.ReleaseReleasedDeprecatedRetiredtoday
ModelLifecycleChronology 2024 → todayReleasedStatements on work productsTested
Gemini 4 Argon GoogleReleased1 statement not verified—
Claude Haiku 5.5 AnthropicReleasedevidenced in 3 products—
Mistral Large 4 Mistral AIPreviewPreview since no statement recorded—
GPT-6.1 Sol OpenAIReleasedevidenced in 3 products—
Claude Sonnet 5.5 AnthropicReleasedevidenced in 5 products—
Claude Opus 5.5 AnthropicReleasedevidenced in 5 products—
GPT-6 Luna OpenAIReleasedevidenced in 4 products—
GPT-6 Sol OpenAIReleasedevidenced in 3 products—
Grok 4.7 SpaceXAIReleasedevidenced in 1 product—
DeepSeek-V4.1-Flash DeepSeekReleasedno statement recorded—
GPT-6 Astra OpenAIReleasedevidenced in 5 products—
Gemini 3.8 Flash GoogleReleasedevidenced in 2 products—
Claude Fable 5.1 AnthropicReleasedevidenced in 7 products—
Claude Mythos 5.1 AnthropicReleasedno statement recorded—

Show all 169 registry entries

No rankingOrders are facts you choose (date, name, provider, evidenced products). agi.jetzt does not rank models.

Watchlist

Models, model families and names beyond the frontier models that we follow. Not a ranking, not part of AGI progress, and no statement about availability.

  • Canonical modelA model in our registry, with what the registry records.
  • Model familyA family a provider names on its own page. Not a single model.
  • Unresolved announced nameA name a provider uses for which our registry has no model.

DeepSeek

  • Model familyDeepSeek V4DeepSeek’s change log lists the generations V4 and V4.1.Source: api-docs.deepseek.com · checked on 11 Oct 2026
  • Canonical modelDeepSeek-V4.1-FlashReleased · 10 Sept 2026
  • Canonical modelDeepSeek-V4-ProReleased · preview from 24 Apr 2026, generally available from 13 Aug 2026
  • Unresolved announced nameDeepSeek V4.1-ProNamed by DeepSeek as an upcoming model, without a date or a page of its own.Source: deepseek.com · checked on 11 Oct 2026

Qwen (Alibaba)

  • Model familyQwen3.8Alibaba offers Qwen3.8 as hosted models and as open weights.Source: github.com · checked on 11 Oct 2026
  • Canonical modelQwen3.8-MaxReleased · 2 Aug 2026
  • Canonical modelQwen3.8-27BReleased · 14 Aug 2026
  • Unresolved announced nameQwen4Named by Qwen as a coming generation; no model of this name has been released.Source: huggingface.co · checked on 11 Oct 2026

Moonshot AI

  • Model familyKimiMoonshot AI offers its models under the name Kimi.Source: platform.kimi.ai · checked on 11 Oct 2026
  • Canonical modelKimi K3Released · date not recorded
  • Canonical modelKimi K2.6Released · date not recorded
  • Canonical modelKimi K2.5Retired

Meta · Muse

  • Model familyMuseMeta introduced Muse Spark on 8 April 2026 as the first model of a new series.Source: about.fb.com · checked on 11 Oct 2026
  • Canonical modelMuse Spark 1.3Released · date not recorded
  • Canonical modelMuse SparkReleased · 8 Apr 2026

Meta · Llama

  • Model familyLlamaMeta’s model overview lists Llama 4 of 5 April 2025 as its newest entry.Source: github.com · checked on 11 Oct 2026
  • Canonical modelLlama 4 MaverickReleased · 5 Apr 2025
  • Canonical modelLlama 4 ScoutReleased · 5 Apr 2025

Mistral AI

Z.ai

  • Model familyGLMZ.ai’s release notes list GLM-5.3 with the date 18 August 2026.Source: docs.z.ai · checked on 11 Oct 2026

Baidu

OpenAI

  • Unresolved announced nameGPT-5.5-CyberNamed so by OpenAI; not in the registry.Source: openai.com · checked on 11 Oct 2026

SpaceXAI

  • Unresolved announced nameGrok 4.1Introduced by the provider on 17 November 2025; not in the registry.Source: x.ai · checked on 11 Oct 2026