Between 1 and 4 September 2026, four frontier laboratories released new flagship models. The compressed timing is the story. Release windows that used to be spaced across quarters now land inside a single week, and the pricing and access decisions announced alongside them say more about the competitive position than the benchmark tables do.
Anthropic opened on 1 September with Claude Fable 5.1, available through the Claude API, AWS, Google Cloud and Microsoft Azure, positioned for long running, high stakes tasks that span hours and multiple applications. Base pricing was unchanged, with cache read prices cut by 75 percent, a move aimed squarely at agent workloads that re-read large contexts. Alongside it, Anthropic released Claude Mythos 5.1, described as the same underlying model with lighter safety constraints, available only to vetted organisations through a trusted access programme.
Google followed on 2 September with Gemini 3.8 Flash on Google AI Studio and the Gemini Enterprise Agent Platform, priced at 0.75 and 3.75 dollars per million input and output tokens through 31 December 2026, rising to 1.50 and 7.50 afterward. Google reported a 90.8 percent score on Terminal-Bench 2.1 and emphasised software engineering and agentic work. A separate variant, Gemini 3.8 Flash Cyber, aimed at finding software vulnerabilities and generating patches, was restricted to trusted defenders through Google's Fairwind programme.
Meta Superintelligence Labs published Muse Spark 1.3 on the same day, through Muse Code and the Meta Model API, at 1.25 and 4.25 dollars per million tokens with a contributor rate of 0.10 and 0.20. Meta reported roughly 20 percent fewer tool calls and 25 percent fewer tokens than its predecessor, with multimodal input and a context window above one million tokens.
OpenAI released GPT-6 Astra as a limited preview on 3 September, with broad availability on 4 September across ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API, Microsoft Azure and AWS Bedrock. Pricing was set at 10 and 50 dollars per million input and output tokens, with long context at 20 and 75 and a regional surcharge of 10 percent. OpenAI reported 97.6 percent on FrontierMath Tier 4, 99.9 percent on ARC-AGI-3 and 100 percent on ExploitBench, and described it as the first OpenAI model to reach the Critical threshold under its own risk framework. Cybersecurity capability was gated behind a Daybreak Access programme.
All of these performance numbers are laboratory reported figures drawn from launch materials, not independent evaluations, and should be read that way.
Two structural signals are worth separating from the marketing. First, three of the four launches carried an access tier that is not open: trusted access at Anthropic, Fairwind at Google, Daybreak at OpenAI. The industry has converged on gating offensive security capability behind vetted programmes rather than withholding it. Second, pricing is now the primary differentiator at the low end, with Google and Meta an order of magnitude below OpenAI per token, while the top of the market has moved toward long horizon autonomous work where the relevant cost is per completed task rather than per token.
As of September 2026, published frontier rankings place Claude Opus 5, GPT-6 Astra and Claude Fable 5 at the top of the field.