Three frontier models in 72 hours: what the September 2026 launches cost

Three frontier launches landed in seventy-two hours. Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. Google made Gemini 3.8 Flash generally available on 2 September. OpenAI announced GPT-6 Astra on 3 September. If you run AI in production, your model menu changed three times in one week — and at least one of those changes moves your unit costs whether you switch models or not.

This post is the pricing read, not the benchmark read. What each model lists per million tokens, what the fine print does to that number, and the four things worth doing on your own bill before the end of the month.

The three launches, side by side

Two of those rows are the same price to the cent. That is the story of this week: the flagship tier has a ceiling, both leaders are sitting on it, and the competition has moved somewhere less visible than the sticker.

Claude Fable 5.1 and Mythos 5.1: the price held, the cache moved

Fable 5.1 and Mythos 5.1 are the same underlying model at two safeguard levels. Fable 5.1 is generally available on the Anthropic API, AWS, Google Cloud, Azure, and Claude.ai. Mythos 5.1 carries more permissive safeguards for defensive cybersecurity and advanced biology work and is restricted to vetted organizations through a trusted-access program, currently US-only. Same rate card either way: $10.00 input, $50.00 output, 1M context, 128K max output.

The change that matters is the cache. On every other Claude model a cache read costs 0.1x the input rate. On Fable 5.1 and Mythos 5.1 it is 0.025x — $0.25 per million against Fable 5's $1.00, a 75% cut. Anthropic puts the effect at roughly 25% lower cost on a typical workload versus Fable 5. Cache writes are unchanged at 1.25x for the 5-minute TTL and 2x for the 1-hour, and batch remains 50% off both sides.

That is a discount you only collect if your prompts are actually structured for reuse, which is why cache-hit rate is the number to instrument before you argue about which model is cheaper.

Gemini 3.8 Flash: cheap now, scheduled to double

Gemini 3.8 Flash is Google's workhorse tier, positioned for long-horizon software engineering and agent loops rather than the flagship bracket. It lists $0.75 input and $3.75 output — including thinking tokens — with cached input at $0.075. A specialized 3.8 Flash Cyber variant ships only through Google's Fairwind program for approved defenders.

Read the date on that price. Those are introductory rates through 31 December 2026. On 1 January 2027 input goes to $1.50, output to $7.50, and cached input to $0.15 — every line doubles. There is also a context-cache storage charge that is easy to miss: $0.50 per million tokens per hour, rising to $1.00 in January. If you are sizing a 2027 budget on a Gemini 3.8 Flash workload, model the post-promotional rate, not the one on the page today.

GPT-6 Astra: matching Anthropic to the cent, if you can get it

GPT-6 Astra lists $10.00 input, $50.00 output, $1.00 cached input, a 1,050,000-token context window, 128K max output, and an April 2026 knowledge cutoff. OpenAI also offers a fast mode at twice the price for twice the speed. Access is staged: limited organizations first, with ChatGPT tiers, the API, Azure, and Bedrock following over the days after launch.

So treat the Astra row as a price you can plan against rather than one you can necessarily buy this week. If your procurement cycle assumes availability, confirm your own access before the number goes in a forecast.

What actually changed: the $10 / $50 ceiling

Six months ago the flagship tier spread across a wide band. Today the two frontier models from the two largest Western labs list identical input and output prices. When sticker prices converge, the cost difference between vendors stops living on the pricing page and moves into three places: cache economics, long-context surcharges, and tokenizer differences that change how many tokens the same text becomes.

The cache line is the biggest of the three right now. Take a retrieval-heavy agent running a 90% cached prefix. Per million input tokens, Fable 5.1 bills 900K cache reads at $0.25 plus 100K fresh at $10.00 — $1.23. GPT-6 Astra bills the same shape at $1.00 and $10.00 — $1.90. Output is identical at $50.00 either way. Same list price, 35% apart on the input side, decided entirely by a rate neither headline mentions. (One-time cache writes are excluded, and a workload with a low hit rate collects none of this.)

The tokenizer caveat cuts the other way and is worth stating plainly: per-token prices are only comparable within a tokenizer generation. Two models at $10.00 per million tokens are not necessarily $10.00 per million words of your traffic. Compare on cost per completed task where you can, not on rate cards alone.

Four things to do on your own bill this month

• ◦Know which model every workload is actually on. Model churn at this pace means a service pinned to a deprecated snapshot, or silently upgraded by a default, is a cost change nobody authored. Per-model spend is the only view that catches it.
• ◦Instrument cache-hit rate per feature. With Fable 5.1 cache reads at 0.025x input, the gap between a 40% and a 90% hit rate is now a larger cost lever than the choice between vendors at the top of the market.
• ◦Calendar the 1 January 2027 Gemini increase. A promotional rate that doubles on a known date belongs in the forecast today, not in January's variance explanation.
• ◦Alert on cost per request per model, not on the monthly total. A model swap or a prompt that grew after a deploy shows up in cost per call within hours; it shows up in a monthly budget after the invoice is written.

None of that requires switching models. It requires knowing what you spend by model, by workspace, and by token type, in the same ledger as the cloud infrastructure the same features run on. That is the job CloudQuell does for Anthropic, OpenAI, and Snowflake spend alongside AWS, Azure, and GCP — allocated to the same cost centers, watched by the same anomaly baselines, at a flat monthly price and free under $10K/month of tracked spend.

What new AI models launched in September 2026?

Three frontier releases in the first three days of the month: Claude Fable 5.1 and Claude Mythos 5.1 from Anthropic on 1 September, Gemini 3.8 Flash from Google on 2 September, and GPT-6 Astra from OpenAI on 3 September. Fable 5.1 and Astra both list $10.00 input and $50.00 output per million tokens; Gemini 3.8 Flash is a workhorse-tier model at $0.75 and $3.75 on introductory pricing.

How much does GPT-6 Astra cost?

$10.00 per million input tokens, $50.00 per million output tokens, and $1.00 per million cached input tokens, with a 1,050,000-token context window and 128K max output. A fast mode is available at twice the price for twice the speed. Access was staged at launch — limited organizations first, then broader API and cloud-marketplace availability — so confirm your own access before planning against the rate.

Is Claude Fable 5.1 cheaper than Fable 5?

The list prices are identical — $10.00 input and $50.00 output — but cache reads drop from $1.00 to $0.25 per million tokens, a 75% cut, because Fable 5.1 and Mythos 5.1 price cache hits at 0.025x input instead of the usual 0.1x. Anthropic estimates roughly 25% lower cost on a typical workload. How much you actually save is a function of your cache-hit rate: a workload that never reuses a prefix saves nothing.

When does Gemini 3.8 Flash pricing go up?

On 1 January 2027. The introductory rates of $0.75 input, $3.75 output, and $0.075 cached input run through 31 December 2026, then double to $1.50, $7.50, and $0.15. Context-cache storage moves from $0.50 to $1.00 per million tokens per hour on the same date. Any 2027 forecast built on the current numbers will be half the real figure.