Blog›Benchmarks

GPT-6 Astra and Claude Fable 5.1: Frontier Launch Week, on the Record

Anthropic, Google and OpenAI all shipped in 72 hours. What launched, what our daily benchmark snapshots already show, and why tiered access is the new frontier pattern.

In the first three days of September 2026, the three biggest AI labs all shipped: Anthropic released Claude Fable 5.1 and Mythos 5.1 on Monday, Google pushed out Gemini 3.8 Flash on Tuesday, and OpenAI announced GPT-6 Astra on Wednesday. Here is what actually launched, what our daily benchmark snapshots already show, and what to watch on the leaderboards next.

What launched

Claude Fable 5.1 and Mythos 5.1 (September 1). Anthropic calls them its most advanced models for coding and knowledge work, with stronger results than Fable 5 at lower cost for many workloads. The two names are one underlying model on two tracks: Fable ships broadly with additional safeguards on dual-use capabilities, while Mythos is offered without them only to vetted organizations (currently a set of US cyberdefenders and life-science labs). Background on the tier in our wiki: Claude Fable 5 and Mythos 5.

GPT-6 Astra (September 3). OpenAI's new flagship, described as the product of "years of research and big bets", rolling out in phases to ChatGPT paid plans, the API and AWS. The headline is not just capability: Astra is the first OpenAI model to reach the company's internal "Critical" cybersecurity threshold, and OpenAI is limiting those capabilities to companies in its Daybreak security program first. That mirrors Anthropic's Fable/Mythos split and makes tiered access the de facto frontier-launch pattern. Wiki page: GPT-6 Astra.

What our snapshots already show

Wikiprompt takes a daily snapshot of the major leaderboards, so launch weeks leave a paper trail:

Fable 5.1 is already listed on OpenRouter in our September 3 snapshot, two days after launch, including a batch pricing tier.

Claude Fable 5 leads the web-development arena on LMArena in today's snapshot with an Elo of 1628 overall (1663 on HTML tasks), ahead of every other model we track there.

On the general text arena, today's top is a tight Anthropic cluster: Opus 5 Max (1505), Opus 5 High (1504) and Opus 4.6 High (1503), with Fable 5 at 1494 alongside Gemini 3.7 Flash High (1491).

GPT-6 Astra is not on the community arenas yet. Phased rollouts typically appear within days; when it lands, the daily snapshot will catch it and the change will be visible at benchmarks history.

Why the access tiers matter

The interesting shift this week is not a benchmark number. Both labs now publicly ship their strongest capabilities behind vetting: Mythos to approved organizations, Astra's cyber capabilities to the Daybreak cohort. For prompt writers this means the model you can buy and the model that tops internal evals are no longer always the same artifact, and community leaderboards (which test the public variants) are the honest measure of what you can actually use.

Watch it move

Today's frozen photo: benchmark snapshot for September 3, 2026

The text arena, updated daily: text leaderboard

Where Fable 5 currently leads: web development leaderboard

Every ranking change on the record: benchmarks history

We will keep snapshotting daily; when Astra and Fable 5.1 hit the arenas, the deltas will be right there.