The frontier lab landscape, September 2026

The week that reset the table

The first week of September 2026 saw four flagship-class releases in five days: Claude Fable 5.1 (September 1), Gemini 3.8 Flash and Meta's Muse Spark 1.3 (September 2), and GPT-6 Astra (September 3). World Labs announced Atlas the same week. Any ranking written before that week is stale, and any written this week will be stale soon. This hub records who is at the frontier and what each lab's distinguishing bet is, not who is ahead by two benchmark points.

The labs

Lab	Flagship (date)	Weights	Distinguishing bet	Note
Anthropic	Claude Fable 5.1 / Mythos 5.1 (2026-09-01)	Closed	One model, two safeguard tiers; constitution-driven post-training; interpretability in the release loop	Anthropic and the Claude modelsNoteSeedAnthropic and the Claude modelsAnthropic's current flagship is Claude Fable 5.1 (released 2026-09-01), shipped as one set of weights under two safeguard configurations. The lab's training story runs from constitutional AI in 2022 to a rewritten constitution in 2026, with interpretability tools now used in pre-release audits.
OpenAI	GPT-6 Astra (2026-09-03)	Closed	Computer use and agent harness; first self-rated Critical cyber model	OpenAI and the GPT lineageNoteSeedOpenAI and the GPT lineageOpenAI's current flagship is GPT-6 Astra (released 2026-09-03), the first model it has rated Critical for cybersecurity. The o-series reasoning line has been folded into the GPT numbering, and the product emphasis has shifted to computer use and the Codex agent harness.
Google DeepMind	Gemini 3.8 Flash (2026-09-02); 3.1 Pro (Feb 2026)	Closed	Native multimodality; 1M context everywhere; owned TPU stack	Google DeepMind and the Gemini familyNoteSeedGoogle DeepMind and the Gemini familyGoogle DeepMind's newest release is Gemini 3.8 Flash (2026-09-02); its top reasoning model is still Gemini 3.1 Pro from February 2026, with Gemini 3.5 Pro unreleased and Gemini 4 in pretraining. Distinctive features are native multimodality, a 1M-token context as standard, and a vertically owned TPU stack.
xAI / SpaceXAI	Grok 4.6 (2026-08-12)	Closed	Owned gigawatt-scale cluster; Cursor, X, and SpaceX data	xAI (SpaceXAI) and the Grok modelsNoteSeedxAI (SpaceXAI) and the Grok modelsxAI, now a SpaceX unit branded SpaceXAI, ships Grok 4.6 (released 2026-08-12) as its flagship. Its documented differences are a very large owned GPU cluster, heavy use of agentic RL and model-generated SFT data, and privileged access to Cursor, X, and SpaceX data. Several widely repeated claims are founder statements, not documentation.
Meta Superintelligence Labs	Muse Spark 1.3 (2026-09-02)	Closed (open release promised, undated)	Cost per task at frontier-adjacent quality	this note
DeepSeek, Moonshot, Alibaba, Zhipu, MiniMax	DeepSeek V4, Kimi K3, Qwen3.8, GLM-5.3, MiniMax M3	Open	Sparse MoE, long-context efficiency, published architecture details	Chinese frontier labs: DeepSeek, Moonshot, Alibaba, and othersNoteSeedChinese frontier labs: DeepSeek, Moonshot, Alibaba, and othersAs of September 2026, at least five Chinese labs ship open-weight models within a few months of the closed frontier. The shared pattern is sparse mixture-of-experts at trillion-parameter scale, aggressive long-context efficiency work, and permissive licenses. Cost claims need care.
World Labs	Atlas (2026-09-01, early access)	Closed	Spatial world models, not language	Atlas: World Labs' spatial world modelNoteSeedAtlas: World Labs' spatial world modelWorld Labs announced Atlas on 2026-09-01 as a single multimodal model for camera-controlled video, 3D reconstruction, and simulation. It is in early access only; no paper, weights, or pricing have been published.

One paragraph per lab

Anthropic ships the model that independent aggregators currently rank highest (Artificial Analysis puts Fable 5.1 at 66 on its index), and it does so with a release pattern no one else uses: identical weights sold as a general product and as a gated product for vetted cyber and life-science users. See Anthropic and the Claude modelsNoteSeedAnthropic and the Claude modelsAnthropic's current flagship is Claude Fable 5.1 (released 2026-09-01), shipped as one set of weights under two safeguard configurations. The lab's training story runs from constitutional AI in 2022 to a rewritten constitution in 2026, with interpretability tools now used in pre-release audits..

OpenAI answered two days later with GPT-6 Astra, positioned less as a chat model than as a computer-use agent. Its own safety overview rates it Critical for cybersecurity under the Preparedness Framework, the first time OpenAI has applied that label. See OpenAI and the GPT lineageNoteSeedOpenAI and the GPT lineageOpenAI's current flagship is GPT-6 Astra (released 2026-09-03), the first model it has rated Critical for cybersecurity. The o-series reasoning line has been folded into the GPT numbering, and the product emphasis has shifted to computer use and the Codex agent harness..

Google DeepMind is in an unusual position: shipping the cheapest competitive models on a monthly cadence while its Pro tier has not moved since February and Gemini 4 is still in pretraining. Its structural advantages (TPUs, distribution) have not yet converted into the top benchmark position. See Google DeepMind and the Gemini familyNoteSeedGoogle DeepMind and the Gemini familyGoogle DeepMind's newest release is Gemini 3.8 Flash (2026-09-02); its top reasoning model is still Gemini 3.1 Pro from February 2026, with Gemini 3.5 Pro unreleased and Gemini 4 in pretraining. Distinctive features are native multimodality, a 1M-token context as standard, and a vertically owned TPU stack..

xAI, now inside SpaceX, competes on owned compute and owned data. Its documentation is more specific than most about the mid-training and agentic RL stages. Several headline claims about the next model are founder statements without a model card. See xAI (SpaceXAI) and the Grok modelsNoteSeedxAI (SpaceXAI) and the Grok modelsxAI, now a SpaceX unit branded SpaceXAI, ships Grok 4.6 (released 2026-08-12) as its flagship. Its documented differences are a very large owned GPU cluster, heavy use of agentic RL and model-generated SFT data, and privileged access to Cursor, X, and SpaceX data. Several widely repeated claims are founder statements, not documentation..

Meta has moved from open Llama to closed Muse Spark. Artificial Analysis scores the available xhigh variant at 61, tied with GPT-5.6 Sol and Grok 4.6, at the lowest cost per task of any model at that level. A max variant is in limited preview. Meta says open weights are coming but has given no version, date, or license; treat that as intent, not a release.

Chinese labs publish what the US labs withhold: parameter counts, active-parameter counts, token counts, optimizer choices, attention designs. DeepSeek's own paper places V4-Pro three to six months behind the closed frontier. That gap, and the open weights, define the segment. See Chinese frontier labs: DeepSeek, Moonshot, Alibaba, and othersNoteSeedChinese frontier labs: DeepSeek, Moonshot, Alibaba, and othersAs of September 2026, at least five Chinese labs ship open-weight models within a few months of the closed frontier. The shared pattern is sparse mixture-of-experts at trillion-parameter scale, aggressive long-context efficiency work, and permissive licenses. Cost claims need care..

World Labs is on this map because "frontier" is no longer only about language. Atlas is a different kind of model with a different evaluation problem, discussed in World models: two meanings of one termNoteSeedWorld models: two meanings of one termThe phrase 'world model' now covers two different research programs: a learned dynamics model an agent plans inside, and a generative model of visual or 3D space. This note separates them and explains why the distinction matters..

What this map does not settle

• Benchmarks are mostly vendor-run. Cross-lab numbers in launch posts use different harnesses and effort settings.
• Parameter counts for every closed model are undisclosed; any figure in circulation is an estimate.
• Access tiers now vary by domain (cyber, biology), so "available" needs a qualifier.

Review log

• 2026-09-08: created; sources checked on this date.