The headline price stopped meaning anything this week
Friday, September 4, 2026·8 min read·4 stories
Three flagship models landed in three days, and two of them carry the exact same sticker price: $10 per million in, $50 per million out. That number now tells you almost nothing. What separates the bills is the cache rate, the context cliff, and what the terms of service quietly take from you. Today is four stories about where the real price hides.
No sponsored or affiliate links in this digest — the links below are sources only. The one product mentioned is our own and labelled as such.
Story I
GPT-6 Astra matches Fable 5.1 to the dollar, then charges four times as much for cache
OpenAI shipped GPT-6 Astra yesterday: a 1,050,000-token context window, 128,000 max output tokens, an April 30, 2026 knowledge cutoff, and computer use put at the centre of the pitch. The benchmark sheet is loud — 72.6% on OSWorld V2-Offline, 74.1% on DeepSWE v1.1, 97.6% on FrontierMath Tier 4. Access is currently limited to organisations in OpenAI's Trusted Access and Daybreak programmes, no weights released, and a "Critical" cyber designation means standard API access blocks exploit-discovery work outright rather than pausing for review.
The pricing is where it gets interesting. Standard rates are $10 per million input tokens and $50 per million output — identical, to the dollar, to Claude Fable 5.1, which we wrote about on Tuesday. Cached input is $1.00 per million. Anthropic's is $0.25. That is a 4x gap on the single line item that dominates an agent loop's bill. There is a second catch further down the page: requests above 272,000 input tokens bill at 2x input and 1.5x output. So the million-token window is real, and it costs $20 per million to actually use past the first quarter of it.
For builders
Do the arithmetic on your own traffic before you switch anything. Take a normal day of agent requests and work out what share of your input tokens are cache reads. At 80% cached — ordinary for a loop that re-sends a stable system prompt and tool set — Fable 5.1 blends to about $2.20 per million input and Astra to about $2.80, so the identical sticker price becomes roughly 27% more expensive on the same workload. That is our own arithmetic from the two published rate cards, and yours will differ; run it with your actual ratio. Then check one more number: your p95 input length per request. If it clears 272K, you are on the premium tier and your spreadsheet is wrong by double. Long-context agents should be trimming retrieval to stay under that line, not celebrating the headline window.
Story II
Meta's Muse Spark 1.3 is cheap. The cheap tier is a data licence with a price on it.
Meta released Muse Spark 1.3 on September 2, available in Muse Code and the Meta Model API with a 1M-token context window. The headline claim is efficiency rather than raw capability: roughly 20% fewer tool calls and roughly 25% fewer tokens than Muse Spark 1.2 on the same agentic work. Capability came along too — 75.4 on DeepSWE v1.1, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 72.7, plus 98.1 on long-context retrieval in the 512K–1M band. Standard pricing is $1.25 in, $4.25 out. Weights are closed; Zuckerberg said on X that an open-weights Muse Spark is coming "soon", with no date attached, and the roadmap has carried that entry for a while.
Then there is the contributor tier: $0.10 in, $0.20 out. That is 12x off input and 21x off output, which is the kind of number that ends an evaluation early. Read the terms before it does. Prompts and outputs may be used to improve Meta's products, and the tier is rate-limited to 100 requests per minute against 3,000 on standard. It is not a discount. It is a data-licensing agreement, throttled to a small fraction of production throughput, with a price attached.
For builders
The efficiency claim is the part worth acting on, and almost nobody measures it. Add two counters to your agent today: tool calls per completed task, and total tokens per completed task. Those are the units your bill is actually denominated in — a model that costs 30% more per token and finishes in 25% fewer of them is cheaper, and no price comparison page will tell you that. Once you have the numbers, model swaps become arithmetic instead of vibes. On the contributor tier: 100 RPM makes it usable for offline evals, batch backfills and your own side projects, and unusable for anything customer-facing. If any client data would pass through it, the answer is no, regardless of price — check what your own customer contracts say about third-party training use before you get tempted.
Story III
Fable 5.1 has three breaking changes, and the third one only breaks for accounts made after August 31
We covered the price cut on Tuesday. Here is the part that actually breaks code. Anthropic's migration guide lists three breaking changes for claude-fable-5-1. Forced tool choice is gone: tool_choice of {type:"any"} or {type:"tool"} now returns a 400, and the replacement is auto plus an explicit instruction plus strict: true schemas. Thinking blocks are model-bound: Fable 5.1 reads blocks from Opus 5 and earlier, but none of those models can read Fable 5.1's. And editing earlier turns invalidates thinking — each block is valid only against the exact system prompt, tools and history that preceded it.
The third one has a detail that will bite unevenly. The check is enforced by default only for accounts created on or after August 31, 2026. Older accounts record the mismatch and carry on unless the request opts in via thinking.block_binding.prefix_mismatch_behavior. Anthropic's own documentation spells out the consequence: if you ship a tool or framework that other people run with their own API keys, your key is probably on an old account and your users are on new ones, so they hit the error before you ever see it. Worth noting alongside this, if you are planning a migration on paper: Fable 5.1 and Mythos 5.1 require 30-day data retention and are not available under zero-data-retention arrangements without express authorisation from Anthropic.
For builders
Two commands, ten minutes. First, find out whether your account is enforced: send a request that edits an earlier turn without the beta header — a 400 naming the header means enforcement is on for you. Second, and this is the one that matters if you ship anything others run, test against the check deliberately by setting prefix_mismatch_behavior, which works from any account regardless of age. Then make the history append-only: freeze system and tools at session start, move mid-session changes into role: "system" messages, and stop rebuilding the messages array from a database on every request. That pattern is not just about the 400 — it is the same discipline that keeps your prompt cache warm, which is exactly the bill from Story I. If you are on a ZDR contract, settle that question with your account team before you write a migration ticket, not after.
Story IV
Gemini Notebook drops fixed daily caps for compute metering, and resets every five hours
Google began rolling out compute-based usage limits for Gemini Notebook on September 2, across every subscription tier on web and mobile. The old system was a list you could read: three audio and video overviews and ten reports or quizzes a day on free, up to 5,000 chat queries and 200 overviews a day on Ultra. The new one is a budget that spends at a rate depending on prompt complexity, chat length, number of sources and which features you use, and refreshes every five hours instead of every twenty-four.
Google frames this as flexible, and for a lot of usage it genuinely is — a short factual query no longer costs the same as generating a full slide deck. But something real is lost in the swap. A fixed cap is a number you can design a workflow around. A compute budget is one you can only discover by spending it. This is the second metering restructure we have written up in a fortnight, after Claude Code's weekly limits at the end of August.
For builders
If you run any client-facing process on a consumer AI tier — research briefs, content batches, transcript summaries — stop treating the published quota as a spec you can quote from. Measure instead: run one real job end to end, note where the budget sits before and after, and you have a per-job cost in units the vendor is actually charging. Do that this week while the old intuitions are still fresh enough to compare against. And the structural lesson, which applies well beyond Notebook: anything you sell on a fixed turnaround needs to sit on a metered API where you can see the invoice, not on a subscription tier whose ceiling the vendor can redefine on a Wednesday.