Interactive LLM timeline
The Evolution of LLMs
Explore the evolution of language models, from the Transformer to reasoning systems and agents. The interactive timeline and story view separate release chronology from the dated benchmark comparison below.
The current comparison
Checked 2026-10-14. AA Intelligence Index v4.3.2: MiMo-V2.6-Pro scores 46.3, versus Claude Opus 5.5 at 57.6. The general reasoning gap is 11.3 points within the reviewed cohort. AA Coding Agent Index v1.5: GLM-5.3 scores 53.6, versus Claude Sonnet 5.5 at 68.4. The coding systems gap is 14.8 points within the reviewed cohort. AutomationBench-AA v1.0.6 objective score: GLM-5.3 scores 62.2, versus GPT-6 Astra at 68.5. The workflow automation gap is 6.3 percentage points within the reviewed cohort.
Benchmark methodology, current general results and coding harness results identify the evidence behind this snapshot. Each number measures a specific model/configuration within its benchmark. Calculations use unrounded values. The reviewed cohort is a subset of the publisher’s results, not a claim to cover every model or self-hosted setup.
Reviewed model catalogue
Claude Fable 5.1Anthropic · hosted
Generally available for long-running coding and knowledge work. Safety fallback routes some tasks to Opus models, so the benchmark configuration must travel with the score.
Claude Fable 5.1 source and access details ↗Claude Mythos 5.1Anthropic · restricted
Available to vetted organisations through trusted-access programmes. This restricted cyber and biology model is tracked separately from the generally available ceiling.
Claude Mythos 5.1 source and access details ↗GPT-6 AstraOpenAI · hosted
OpenAI documents Astra for complex reasoning, coding, computer use and professional work. The API supports text-and-image input, text output and tool use through Responses. Product access and advanced cyber permissions remain separate.
GPT-6 Astra source and access details ↗Gemini 3.8 FlashGoogle · hosted
Google's new Flash model targets longer software and enterprise workflows. Its hosted distribution and speed-oriented positioning are separate from open-weight availability.
Gemini 3.8 Flash source and access details ↗Muse Spark 1.3Meta · hosted
Meta now documents public pay-as-you-go access to Spark 1.3 with all reasoning efforts, including max in Standard mode. Contributor mode does not support max. Audio understanding in 1.3 is not fully supported. The earlier restricted-max observation remains archived.
Muse Spark 1.3 source and access details ↗GLM-5.3Z.AI · open-weights
Downloadable weights under a custom GLM-5.3 licence, separate from GLM-5.2 MIT terms. General, coding and workflow results retain their own benchmark versions and configurations.
GLM-5.3 source and access details ↗GLM-5.3-FlashZ.AI · open-weights
Downloadable multimodal Flash weights are verified in this sweep. MIT licensing and a lower cost per benchmark task make this a distinct deployment option from the larger custom-licensed GLM-5.3.
GLM-5.3-Flash source and access details ↗Qwen3.8 2.4T A95BAlibaba · open-weights
The downloadable text checkpoint is distinct from hosted Qwen3.8-Max, whose additional vision, tools and serving features must not be attributed automatically to the weights.
Qwen3.8 2.4T A95B source and access details ↗Qwen3.8 27BAlibaba · open-weights
A smaller downloadable Qwen3.8 model broadens the practical deployment floor. The observed date records this verification, rather than inferring a public launch from repository creation.
Qwen3.8 27B source and access details ↗Qwen3.8-Flash-NextAlibaba · open-weights
An experimental Qwen4 architecture preview combines sparse attention, gated residuals and n-gram embeddings. Its 125B language model activates 6B parameters, with another 51B embedding and 4B multi-token-prediction parameters.
Qwen3.8-Flash-Next source and access details ↗DeepSeek V4 Pro 0813DeepSeek · open-weights
Matching MIT-licensed 0813 weights remain downloadable. DeepSeek says its hosted deepseek-v4-pro alias routes to V4.1-Flash from September 14, so API aliases must not be treated as immutable 0813 checkpoints.
DeepSeek V4 Pro 0813 source and access details ↗DeepSeek V4 Flash Vision ExpDeepSeek · open-weights
An experimental downloadable vision branch is verified. Its availability is recorded without assigning an unverified comparable general or coding score.
DeepSeek V4 Flash Vision Exp source and access details ↗MiniCPM5-2BOpenBMB · open-weights
OpenBMB's compact local model targets tool use, coding and reasoning. It has about 2.52B total parameters, around 1.98B outside embeddings, and native 131,072-token context. Vendor averages are not an AA Intelligence Index score.
MiniCPM5-2B source and access details ↗Agnes 2.5 Pro BetaSapiens AI · hosted
A hosted beta for multimodal reasoning and agentic work. The August launch evaluation used an older Intelligence Index, so its original 49-point result is not inserted into the September v4.3 comparison.
Agnes 2.5 Pro Beta source and access details ↗Gemini 3.8 Flash CyberGoogle · restricted
A specialised model for vulnerability detection and patching, offered through the Fairwind programme. Its restricted security access is distinct from ordinary Gemini 3.8 Flash.
Gemini 3.8 Flash Cyber source and access details ↗Claude Opus 5.5Anthropic · hosted
Hosted frontier model at $4 input and $20 output per million tokens. The current general result uses max effort with production fallback safeguards; a separate Claude Code max-effort coding result is now captured.
Claude Opus 5.5 source and access details ↗GPT-6 SolOpenAI · hosted
The earlier Sol remains recorded separately from GPT-6.1 Sol. Standard short-context API pricing is $2 input and $10 output per million tokens. OpenAI fixed an image-encoding bug on September 25; older benchmark captures remain dated evidence.
GPT-6 Sol source and access details ↗GPT-6 LunaOpenAI · hosted
Efficient hosted reasoning model with text-and-image input at $0.10 input and $0.50 output per million tokens for short contexts. The Decisions API beta added on October 6 returns typed answers; it is a separate API surface, not a new model checkpoint.
GPT-6 Luna source and access details ↗Grok 4.7xAI · hosted
Available through Grok Build, Cursor and the API. The captured independent results use xhigh effort and a named Grok Build coding configuration.
Grok 4.7 source and access details ↗DeepSeek V4.1 FlashDeepSeek · open-weights
Downloadable MIT multimodal weights and the deepseek-flash API. The earlier V4 Flash and Vision Exp routes are retired compatibility aliases; deepseek-v4-pro also routes here from September 14. Historical checkpoint measurements remain distinct.
DeepSeek V4.1 Flash source and access details ↗MiMo-V2.6-ProXiaomi · open-weights
The official Pro-RL checkpoint is downloadable. The independently measured Pro configuration leads the captured open-weight general cohort. Verification date is not a claimed launch date.
MiMo-V2.6-Pro source and access details ↗MiMo-V2.6-Flash-RLXiaomi · open-weights
An efficiency-focused multimodal checkpoint with downloadable weights. No comparable result in the captured general or coding datasets, so it remains unscored.
MiMo-V2.6-Flash-RL source and access details ↗MiMo-V2.6-Distill-Qwen-9BXiaomi · open-weights
A smaller downloadable distilled checkpoint. Its size and availability do not establish parity with the Pro model. No score is inferred.
MiMo-V2.6-Distill-Qwen-9B source and access details ↗Ling-3.0-flash-FinAnt Group · open-weights
A finance-specialised checkpoint for research and tool workflows. Domain evaluations are not substituted for the general Intelligence Index.
Ling-3.0-flash-Fin source and access details ↗Step 5 PreviewStepFun · hosted
Hosted preview verified through Vercel AI Gateway. Downloadable weights have not been verified, so this preview is excluded from the open floor.
Step 5 Preview source and access details ↗Qwen3.8 Max (0902)Alibaba · hosted
The hosted September 2 serving revision is separate from the downloadable Qwen3.8 text checkpoint. Its independent score is assigned only to this API configuration.
Qwen3.8 Max (0902) source and access details ↗Gemini 3.8 LiveGoogle · hosted
A generally available audio-to-audio model for real-time voice through the Live API. It has no comparable text Intelligence Index score in this snapshot.
Gemini 3.8 Live source and access details ↗Gemini 3.8 Live Extended ThinkingGoogle · hosted
A generally available audio-to-audio model for real-time voice through the Live API. It has no comparable text Intelligence Index score in this snapshot.
Gemini 3.8 Live Extended Thinking source and access details ↗GPT-6.1 SolOpenAI · hosted
Hosted coding and professional-work model at $2 input and $10 output per million tokens for prompts up to 272K tokens. OpenAI added Responses API Ultrafast mode on October 8. Processing tier changes do not constitute a new model or a new benchmark score.
GPT-6.1 Sol source and access details ↗Claude Sonnet 5.5Anthropic · hosted
Hosted coding and document model with $2 input and $10 output per million tokens. Anthropic reduced cache-read pricing to $0.10 per million tokens on October 7. General and Claude Code results retain effort and fallback configurations.
Claude Sonnet 5.5 source and access details ↗Gemini 4 ArgonGoogle · hosted
Google opened public access on October 13 through the Gemini API and subscriber products. Pro and Ultra include Gemini app access; Antigravity requires Ultra. The earlier Fairwind-only observation stays archived. The captured general and Antigravity CLI configurations can now enter the ordinary hosted cohort.
Gemini 4 Argon source and access details ↗Ling 3.1 FlashAnt Group · hosted
Hosted through Vercel AI Gateway since September 30, with 560B total and 25B active parameters and a 262K-token context. The free promotion ended October 13. Downloadable weights and a public weight licence remain unverified.
Ling 3.1 Flash source and access details ↗MiniMax-M3.1-Flash-PreviewMiniMax · hosted
A hosted preview for coding, research and visual work. The official guide describes text, image and video input, text output and always-on reasoning with five effort levels. Access depends on the account; no comparable score or public checkpoint is inferred.
MiniMax-M3.1-Flash-Preview source and access details ↗Solar Mini 4Upstage · hosted
Upstage documents a 35B model with 3B active parameters for repetitive agent workloads, available through its API, Solar Chat and OpenRouter. On-premises commercial support does not establish public downloadable weights. The observed date avoids conflating API snapshot and announcement dates.
Solar Mini 4 source and access details ↗OrcaSAQ-2 27BOrcaRouter · open-weights
An Apache-2.0 compressed Qwen3.8-27B checkpoint. The publisher reports roughly 12.3 GB of weights for deployment on a 16 GB GPU. Memory and fidelity results are vendor measurements, not a comparable Intelligence Index score.
OrcaSAQ-2 27B source and access details ↗Ternary Bonsai 2 27BPrism ML · open-weights
Downloadable Apache-2.0 ternary Qwen3.8-27B-derived weights with GGUF and a companion MLX build. The publisher reports an approximately 5.9 GB language checkpoint; specialised runtimes and a separate vision pack matter. No cross-benchmark capability ratio is inferred.
Ternary Bonsai 2 27B source and access details ↗Qwen3.8-Omni-FlashAlibaba · hosted
Alibaba Cloud documents text, image, audio and video input with text output, custom tool calling and adjustable thinking. The hosted endpoint is verified separately from any open-weight Qwen family.
Qwen3.8-Omni-Flash source and access details ↗Gemini 3.8 Flash TTSGoogle · hosted
Generally available text-to-speech model for expressive, long-form voice generation. Audio generation is tracked separately from text reasoning and receives no general benchmark score.
Gemini 3.8 Flash TTS source and access details ↗Gemini 3.8 Flash-Lite TTSGoogle · hosted
Generally available lower-cost text-to-speech option for high-throughput production and voice-agent cascades. Its audio role does not establish a general reasoning score.
Gemini 3.8 Flash-Lite TTS source and access details ↗GPT-Live 1OpenAI · hosted
Generally available full-duplex voice model. Voice sessions can delegate reasoning and tools to a backend model; session billing and backend model usage are separate. Not assigned a text Intelligence Index score.
GPT-Live 1 source and access details ↗GPT-RosalindOpenAI · restricted
Available through trusted access for approved internal life-sciences research. API billing began October 5, 2026. Restricted research access is excluded from the ordinary proprietary ceiling.
GPT-Rosalind source and access details ↗Claude Haiku 5.5Anthropic · hosted
Available through Claude Platform and partner clouds for fast, high-volume work, with adjustable effort. Input/output pricing is $0.10/$0.50 per million tokens up to 100K prompt tokens and $0.50/$2.50 above that threshold. General max and Claude Code max results remain separate configurations.
Claude Haiku 5.5 source and access details ↗Mistral Large 4Mistral AI · hosted
Public API preview with multimodal input. Current documentation specifies 1.05T total and 52B active parameters, superseding the earlier 49B note. Mistral promises weights by the end of October; pending weights cannot set the downloadable floor.
Mistral Large 4 source and access details ↗Reflection BeamReflection AI · restricted
Selective early-access preview for coding, reasoning and agents, with a waitlist. The 501B-total / 23B-active model is announced with Apache-2.0 weights planned later in October. No downloadable checkpoint or comparable score is verified.
Reflection Beam source and access details ↗Kolibri-1Aleph Alpha · open-weights
German and English reasoning model with 78B total and about 3B active parameters. Aleph Alpha dates availability to October 3 and its announcement to October 5. Downloadable weights are verified; vendor domain evaluations are not an independent general score.
Kolibri-1 source and access details ↗Muse WonderMeta · hosted
Hosted multimodal reasoning model available through Meta Model API and Muse Code. The public API model is muse-wonder-1.0. Its max-effort setting does not select Wonder Full, which is a separate subscription model.
Muse Wonder source and access details ↗Muse Wonder FullMeta · hosted
Higher-capability Muse Wonder variant offered through Muse Code Power Usage and Maximum Usage subscriptions. It is not available through the pay-as-you-go Model API. General and Muse Code max results are captured separately from Wonder.
Muse Wonder Full source and access details ↗Muse Glimmer 30BMeta · open-weights
Local multimodal model distilled from Muse Spark, with official bf16, GGUF and device runtime artifacts. Its checkpoint and licence are re-verified in this review. The observed date is not a new launch date.
Muse Glimmer 30B source and access details ↗GPT Image 2.5 SunburstOpenAI · hosted
Hosted image generation and editing model for precise edits through Image and Responses APIs. Image quality settings are separate from text reasoning benchmarks.
GPT Image 2.5 Sunburst source and access details ↗GPT Image 2.5 FlareOpenAI · hosted
Hosted image model for fast everyday generation and editing. No text Intelligence Index score is inferred from image quality.
GPT Image 2.5 Flare source and access details ↗Nano Banana 2.1Google · hosted
Generally available image generation and conversational editing model, gemini-nano-banana-2.1. Google deprecates the older gemini-3.1-flash-image alias without announcing a shutdown date. Image performance stays separate from text benchmarks.
Nano Banana 2.1 source and access details ↗Qwen Image 2.1 TurboAlibaba · open-weights
Downloadable accelerated image-generation checkpoint with a publisher model card and actual weight files. Availability is observed on this review date; it is not represented as a text LLM or given a text benchmark score.
Qwen Image 2.1 Turbo source and access details ↗Nemotron 3.5 LightningNVIDIA · open-weights
Downloadable hybrid Mamba/attention MoE for specialised agent tasks, with 3B active parameters and BF16/NVFP4 builds. This August release was missing from the reviewed catalogue. Its deployment and vendor speed claims are not a comparable general score.
Nemotron 3.5 Lightning source and access details ↗Hy4 PreviewTencent · open-weights
Downloadable flagship preview with 770B backbone parameters and a separate 10B multi-token-prediction layer. The backbone activates 49B parameters. Availability and actual files are verified today without inferring a launch date or independent frontier score.
Hy4 Preview source and access details ↗How to read the numbers
General reasoning uses Intelligence Index v4.3.2. Coding uses Coding Agent Index v1.5 with the named harness and effort; one explicit configuration per model is retained and excluded alternatives are recorded. Workflow automation retains the September 7 AutomationBench-AA v1.0.6 objective dataset, re-fetched October 14, not a new evaluation of October releases. Objective completion is not whole-workflow success. Compare only within a domain and benchmark version. Hosted results do not guarantee identical self-hosted performance. Scores retain source precision and display one decimal. The October 14 capture reads the publisher’s server-rendered tables: 27 general models and 39 coding configurations, including explicit exclusions. Coverage is a reviewed subset, not the full landscape. Earlier snapshots stay frozen. Missing scores remain absent. Argon and Spark max now have verified ordinary hosted access; Mythos, Flash Cyber and Beam remain restricted. Pending Mistral and Beam weights cannot set the downloadable floor.
The floor means downloadable capability, not a promise that the model will fit on a laptop or that its licence is unrestricted. Restricted-access models, including Mythos and Reflection Beam, are excluded from the ordinary proprietary ceiling. Missing comparable scores remain unscored.