Four labs, 72 hours: what the September 2026 frontier wave actually changed

Between the morning of 1 September and the afternoon of 3 September 2026, four frontier labs shipped a flagship or flagship-adjacent model. Anthropic on the Monday, Google and Meta on the Tuesday, OpenAI on the Wednesday. Seventy-two hours is not a coincidence — nobody schedules against a rival by accident — and it makes the week unusually useful, because the four releases can be read against one another at a single moment in the market rather than months apart.

What actually shipped, in order

Monday 1 September — Anthropic, Claude Fable 5.1. Base price unchanged at $10 per million input tokens and $50 output; cache reads cut 75%, from $1.00 to $0.25. 1M-token context, 128K maximum output. The headline capability number is 52.6% on Terminal-Bench-Science 0.1, against 24.7% for the prior Fable 5. A companion, Mythos 5.1, is the same model under a restricted-access programme for vetted cybersecurity and life-sciences organisations.

Tuesday 2 September — Google, Gemini 3.8 Flash and 3.8 Flash Cyber. Google's third Flash in six weeks. $0.75 input / $3.75 output — but only until 31 December 2026; from 1 January 2027 the price doubles to $1.50/$7.50. Reported 90.8% on Terminal-Bench 2.1, up from 81.6% for 3.7 Flash. The Cyber variant is not generally available at all: it runs through a new gated programme called Fairwind.

Tuesday 2 September — Meta, Muse Spark 1.3. Pricing held at $1.25/$4.25 with a 1M context, plus a new Contributor tier at $0.10/$0.20 — around 12× cheaper on input and 21× on output — in exchange for Meta training on the user's prompts and responses.

Wednesday 3 September — OpenAI, GPT-6 Astra. $10 input / $50 output, 2.5× the rate of GPT-5.6 Sol. 1,050,000-token context, 128K output. Rolled out first to vetted enterprise customers through a gated programme called Daybreak, ahead of ChatGPT plans and the API.

The three things the week changed

1. The premium tier came back. For most of 2026 the story was compression: frontier capability at ever-lower prices, with $6–10 per million output tokens becoming the default band. Astra breaks that pattern deliberately, at 2.5× its predecessor's rate. Fable 5.1 holds the same $10/$50 line. Two of the four labs are now explicitly selling a top tier that costs a multiple of the workhorse below it — which is a change from "the newest model is the cheap one" to "the newest model is the expensive one, and the previous generation is what got cheap."

2. Pricing got a shelf life. Google's 3.8 Flash price is an introductory rate with a printed expiry; it doubles on 1 January 2027. Meta's Contributor tier is a discount with a non-monetary cost attached. Anthropic's change is to cache reads, not base rate — meaning the effective price now depends heavily on how much of a workload is repeated context. None of the four launch-day numbers is a stable "price of the model" in the way a rate card used to be.

3. The most capable variants are being gated, on purpose. Google's Flash Cyber, OpenAI's Daybreak programme and Anthropic's Mythos 5.1 all describe the same shape: a model or variant whose full capability is available only to vetted organisations, with a public version that is deliberately more constrained. Three labs, one week, the same structure. That is a market convention forming, not a one-off.

What a buyer should actually do with this

Do not compare these four on a single benchmark — they were not marketed on one. Astra led with computer use and ARC-AGI-3, Fable with Terminal-Bench- Science, 3.8 Flash with Terminal-Bench 2.1, Muse with price. The honest comparison is per-task, on your own evaluation, which the wide price spread now makes worth doing: a $10/$50 model has to be clearly better than a $0.75/$3.75 one on your work to justify a 13× output-cost difference.

Treat every price above as dated. Google's expires; Meta's has strings; the others have moved before and will again. Re-check at the point of use, and build in the assumption that the number changes.

And ask, for any "cyber-capable" or "most capable" variant, whether the version you can actually buy is the one the benchmarks were run on. This week, for three of the four labs, the answer was not always yes.

---

Sources

• Anthropic — VentureBeat · MarkTechPost
• Google — blog.google, Gemini 3.8 Flash and 3.8 Flash Cyber · Help Net Security
• Meta — Crypto Briefing · OpenRouter
• OpenAI — VentureBeat · CNBC · ARC Prize