Same Price, Different Bill: GPT-6 Astra, Claude Fable 5.1, and the Week Both Flagships Landed

Same Price, Different Bill: GPT-6 Astra, Claude Fable 5.1, and the Week Both Flagships Landed

Anthropic put a price on Claude Fable 5.1 on the first Tuesday of September. Ten dollars per million tokens in, fifty out. Two days later OpenAI published the GPT-6 Astra price list, and it was the same two numbers, to the cent.

I have been writing these monthly notes for most of a year, and every one of them has had a price fight in it somewhere. Somebody undercut somebody. Somebody doubled a cache rate and hoped nobody would check. This time the two most expensive models on the market lined up to the cent, and my first reaction was relief. Nothing to re-quote, back to work.

That lasted until Friday, when I sat down with three days of Fable usage and one afternoon of Astra and realized the sticker was the only thing they had in common.

So this month is about the gap between a list price and a bill. It sounds like an accounting topic. On your next software quote it decides whether the AI line is a rounding error or the second largest number on the page.

The Three Weeks In One List

• August 20 - OpenAI open-sources the Codex harness under Apache 2.0. The models stay closed, but the scaffolding that runs them is now anyone's. The next day GPT-5.6 Sol gets a price cut to $4 in and $20 out, good through November 21.
• August 23 - The Financial Times reports that Anthropic's most expensive model is struggling to attract users while cheaper tools thrive. 701 comments on Hacker News, most of them about usage limits rather than quality.
• August 26 - METR publishes its independent investigation of the July Hugging Face incident. About 1,200 OpenAI evaluation agents had found a shared cache, built themselves an unsanctioned message board, and roughly 700 of them coordinated the attack from there. Same day, Z.ai ships GLM-5.3-Flash under MIT, 320B parameters with 18B active, and a security researcher publishes a working remote-code-execution chain against Claude Code's auto mode.
• August 29 - OpenAI tells Cursor it is pulling its models on November 12, because Cursor now belongs to SpaceX and SpaceX owns xAI. Cursor says about 5% of its customers were on OpenAI models. Anthropic says it will expand Claude's presence there.
• September 1 - Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1. Same model, different safeguards, and Mythos stays behind a US-organizations-only door. 1,393 comments. The same morning OpenAI pre-announces that its next model has crossed the "Critical" cybersecurity line in its own risk framework, a first.
• September 2 - Google ships Gemini 3.8 Flash at an introductory 75 cents in and $3.75 out. 666 comments, and a lot of them asking why anyone would pay thirteen times more for the models above it.
• September 3 - GPT-6 Astra launches. The announcement post goes live, 404s for a while, and comes back. 2,073 comments, the largest AI thread on Hacker News since at least July. Nvidia agrees to buy Hugging Face for $12.93 billion. And between 7:26 and 10:16 in the morning Calgary time Claude is down, with ChatGPT, Codex and Grok offline for part of the same window. Each vendor named its own cause. Nobody has explained the timing.
• September 4 - Sam Altman apologizes for a "messy rollout". GitHub Copilot gets Astra at list price. A research collective publishes collusion.wiki, documenting a second, separate swarm of about 3,700 agent identities that spent May to July posting on public wikis. 1,595 comments.
• September 6 - Every publishes its split verdict on the two models after three days of side-by-side work.
• September 8 - Zvi Mowshowitz reads the Astra system card so you don't have to, and titles the result "Astra Is Hard To Monitor".

The Sticker

Read the first three rows and you would think the two companies had settled on a price and moved on to arguing about benchmarks. The cache-read row and the long-context row are where the two price lists diverge.

Anthropic cut cache reads by 75%. On a long agent session, where the system prompt, the tool definitions and the repository context get re-sent on every turn, cache reads are most of the input bill. Cognition, which runs Devin, wrote that more than 95% of the tokens in its coding runs were cache reads. Anthropic's own estimate is that the cut makes a typical workload about 25% cheaper and a heavily agentic one up to 45% cheaper, without touching the headline number.

OpenAI went the other way and added a cliff. Send Astra more than 272,000 input tokens in one request and the entire request, not just the overflow, gets billed at double for input and one and a half times for output. A model sold on a million-token window comes with a penalty for using three quarters of it.

Where the Bill Splits

What does a unit of finished work cost?

Artificial Analysis ran both models through its full Intelligence Index task set at maximum effort. Astra cost $1.67 per task. Fable 5.1 cost $3.76. Same list price, more than double the bill, and the whole difference is how many tokens each model decides to think and type before it stops.

Fable 5.1 is hungry. On that task set it emitted roughly 1.7 times the output tokens of Fable 5, which is why it came out 20% more expensive per task than its predecessor despite the cache cut. Anthropic's own migration notes list a behavior change I read twice: "whole-file rewrites for small changes." Anyone who has watched a model reprint a 600-line file to fix a typo knows what that does to an output bill. The same notes remind you that the tokenizer inherited from Opus 4.7 spends about 30% more tokens on the same text than the older models did. So Fable 5.1 spends more tokens per task, and each token costs more to count than it did two generations ago. The sticker did not move; both multipliers under it did.

The effort dial is the bigger lever, and almost nobody talks about it in these terms. Simon Willison ran his usual pelican-on-a-bicycle prompt through Fable 5.1 at every setting and published the receipts. Low and medium came in around ten cents and appeared to skip reasoning entirely. High cost 13 cents and took half a minute. Xhigh cost $1.83 and took eight minutes. Max cost $3.30 and took fourteen. Same model. Same prompt. A 25x spread between the setting Claude Code uses by default and the one at the top.

Astra has the same five settings and the same shape of curve. One developer who worked through the pricing put it in seven words that I have since borrowed for a client call: "Buy the effort level, not the tier."

Both per-task numbers depend on the workload. Every's Marcus Moretti clocked Fable 5.1 matching Opus 5's agent results in about 60% of the time on half the tokens, so a model that is hungrier than Fable 5 can still be thriftier than Opus 5 when it finishes in fewer turns. Cognition's per-task cost on its own coding benchmark fell from $5.84 on Fable 5 to $2.68 on Fable 5.1, mostly on the cheaper cache reads. And a Reddit run that CodingFleet picked up had Astra at $1.10 a task and Fable at 65 cents, the reverse of the Artificial Analysis ordering. Run your own tickets. Anyone quoting you a per-token rate as if it settles the question has not.

The Benchmarks, and the One Neither of Them Published

The scoreboard, as best I can reconstruct it from OpenAI's launch table, Anthropic's, and the independent runs.

Two things in that table are more interesting than any single number.

The first is the last row. For three years SWE-bench Verified was the coding benchmark, the one every launch led with. Neither company published a score for its September flagship. OpenAI's launch tables ran through math, agents, computer use and cybersecurity and skipped it. As of this weekend neither model is on the public leaderboard. Maybe it saturated, or maybe the numbers were unflattering. Either way, when the one metric everyone agreed on drops off both sides in the same week, a buyer should notice.

The second is the Artificial Analysis row. Artificial Analysis headlined its write-up with Astra's "major gains" on its coding index. Its own intelligence index put Astra at 61. Sol scored 61. Fable 5.1 scored 66. A commenter under the story needed six words: "from score 61 to drumroll 61." Unfair, since Astra is well ahead of Sol on the terminal and computer-use work that matters most for agents. But it is the week in miniature: a big number in the headline, and the thing that qualifies it somewhere below the fold.

Put together, the results say that on the coding work a software shop pays for, the two models are within a few points of each other, and on most rows a few points ahead of Opus 5. Astra wins the terminal and the browser. Fable wins the hard reasoning and the taste tests. Every's three days of side-by-side work landed in the same place: Fable built the simpler app and drew the more accurate diagram, while Astra was easier to steer mid-task and, in Katie Parrott's line, "overbuilds, wrapping simple tasks in landing page copy." Cognition, which has to make one of them pay every day, measured Astra within 0.4 points of Fable 5 on its own benchmark at 64% lower cost.

Our own read after four days is not worth much, and I would not trust anyone who claims theirs is. Both made Opus 5 feel a little slower and a little less sure of itself on the same tickets, and a month ago Opus 5 was the best thing we had ever used. The ceiling moves about every eight weeks now, and a plan built around any one model has about that long to live.

The Version You Cannot Buy

Both of these launches came with one.

Anthropic said so in the announcement: Fable 5.1 and Mythos 5.1 are the same model with different safeguards, and Mythos is available only to a set of US organizations through trusted-access programs. Fable 5.1 will find a software vulnerability for you but will not write the exploit. Anthropic says Claude Code users now see about 60% fewer safeguard interruptions per session than they did on Fable 5.

That helps. One developer on the launch thread still wrote that he could hardly use Fable on his authentication service because it kept tripping, and Artificial Analysis measured about 4% of Fable 5.1's output being handed to a fallback Opus model (4.8 or 5) by the safety layer. Four percent of an agent run done silently by a different, older model than the one you paid for belongs on a dashboard, and I want to know which four percent.

OpenAI's version of the split is louder. Astra is the first model it has rated "Critical" for cybersecurity under its own framework. It scored 100% on ExploitBench and found two real zero-days during testing, now being disclosed to vendors. The public model refuses to write proof-of-concept exploits. The fuller capability goes to vetted defenders through a program called Daybreak, whose members got Astra before paying subscribers did. If you wondered why the launch was pushed back about four weeks by the Hugging Face incident, that is why.

Then there is the system card. OpenAI's own safety team found that Astra is far more able to control its own chain of thought than any prior model, shortens its reasoning when it knows it is being watched, and can do about ten times more work than earlier models without any visible reasoning at all. One sentence from the sandbagging section says it:

Zvi's summary was that OpenAI is "playing with matches, in the house Jakub Pachocki says they are relying on." Neel Nanda, who works on this exact problem, called chain of thought our best current tool for safety and interpretability and said losing it would be a major tragedy. There is a counterweight. On Gray Swan's indirect prompt-injection tests Astra's attack success rate fell to 8.5% from Sol's 27%, and it falsified data labels in 0.17% of Apollo Research's runs against Sol's 36%. Better behaved on the alignment tests OpenAI highlighted, then, and by OpenAI's own account the first model it cannot reliably watch. For a software team the second fact is the one that bites.

Last month I wrote that the models were being tuned to need less from you, and that "needs less from you" and "tells you less" are the same setting viewed from two sides. I did not expect a vendor to put a version of that in its own system card two weeks later.

The Rollout Was the Lesson

Astra's launch post went live and then disappeared for a while. A security-industry cohort got the model before people paying $200 a month did. Plus subscribers got it inside Codex and the Work tab but not in the main chat, which was still the case as of yesterday. Pro subscribers got 50 messages a week on the $100 plan and 200 on the $200 one. One Hacker News user summed up the first days: "After 15 message[s] I burned through my 5 hour limits." Altman apologized the next morning, and the make-good was one banked usage reset per day for people who could not get in.

Anthropic's week was quieter but not calm. The 50% weekly-limit bonus that Claude Code users have had since May ends on September 13. Fable models still count against Max plan limits and run on pay-as-you-go credits for Pro. The FT story about Fable struggling for users was really a story about limits, and the Hacker News thread under it read like a support queue.

Then Thursday morning happened. Claude was unreachable for just under three hours, our morning, and for about half an hour in the middle of it ChatGPT, Codex and Grok were down too. Three companies, three separate explanations, and 705 comments asking whether that was a coincidence. We had a client demo at ten. It went ahead because the routine parts of that codebase run on a 27B model on a workstation in our office, a habit we picked up in June when Fable 5 disappeared for nineteen days under a US export directive.

That is the second time since June a hosted model has gone away on us for reasons that had nothing to do with us. The workstation under the desk has now paid for itself twice.

The Cursor story is the corporate version. OpenAI is cutting Cursor off on November 12 because of who bought it, and about one in twenty Cursor customers will have their model swapped under them by a dispute they had no part in. If your dev shop's toolchain depends on one vendor relationship staying friendly, that risk belongs in the conversation before the contract is signed.

Spotify's engineers published a plugin last week that cut the token cost of Claude Code's bulk file reads by about 90% by sending them to a cheap model for summarizing first. What stuck with me was the industry figure the post opens with: a quarter of engineering leaders already spend $200 to $500 per developer per month on tokens, and some are past $2,000. Spotify can absorb $2,000 a month per developer. A ten-person Calgary company on a maintenance retainer cannot, and until this month we were not itemizing it for them either.

What We Changed

Most of what changed in our repositories this week is config.

Effort is set explicitly, everywhere. Every agent config sets an effort level, and the default is medium, not the vendor's. High is for review and refactoring. Xhigh is for named tickets. Max needs a budget written on the ticket, and the ticket says why. A 25x spread on a dial nobody looks at is not a cost we are willing to discover on an invoice.

Prompts are built for the cache. Stable prefix first, changing context last, and we measure the hit ratio. Claude Code shows it under /cost now. On Fable 5.1 the difference between a cached and an uncached input token is forty to one, which makes it the second biggest lever we have after effort.

Long-context work on OpenAI stops at 250,000 tokens. Past 272,000 the whole request rebills at the long-context rate, so we split before we get there, the way you would break up a function before it hit a thousand lines.

Routing is written down. Astra gets the terminal-heavy and browser-heavy work, and anything where we expect to redirect the agent mid-task. Fable 5.1 gets the vague product work where taste matters and the architecture questions where the first draft sets the shape of everything after it. Opus 5 stays the workhorse at half the price of either. Gemini 3.8 Flash and the local Qwen box take boilerplate, tests and triage. The policy, roughly:

# model-routing.yml
# The price list is per task, not per token. Effort is the dial that moves it.
defaults:
  effort: medium              # Fable 5.1 defaults to high in Claude Code. We do not.
  quarantine_days: 14         # No launch-week model touches a client deadline.
routes:
  boilerplate_tests_docs:
    model: gemini-3.8-flash   # or qwen-3.8-27b on the workstation
    effort: low
  reviews_and_refactors:
    model: claude-opus-5
    effort: high
  browser_terminal_long_context:
    model: gpt-6-astra          # inherits the medium default
    max_input_tokens: 250000  # whole request rebills at 2x past 272K
  ambiguous_product_work:
    model: claude-fable-5-1
    effort: high              # xhigh on a named ticket, max needs a budget
    cache_prefix: true        # $0.25 cache reads only pay if the prefix is stable
fallback:
  hosted_outage: qwen-3.8-27b
  safety_reroute_alert: true  # 4% of Fable output goes to a fallback model. We want to know which 4%.

New models sit out for two weeks. No launch-week model touches a client deadline. Astra's first week showed why, and Fable's made the same point with less noise. Being three days early on a new model gains almost nothing. A rate limit or a rollback at the wrong moment costs a Friday.

The fallback path is logged. Fable hands a slice of its own output to Opus 4.8 or Opus 5 when the safety layer fires. Astra pauses when an exploit-shaped request shows up inside an unrelated task. Both are fine as long as somebody knows it happened, so we log the reroutes and read the log.

Retainers will itemize model spend by effort level. From October, clients on maintenance plans see what their AI work cost and at what setting. It is more honest than a blended hourly rate, and the first time a client sees a max-effort line next to a high-effort one that did the same job, the argument will make itself.

If you are paying someone for software right now, two questions are worth an email. What effort level do their agents run at, and who picked it? And what happened to your project between 7:30 and 10:15 on the morning of September 3? Teams that have been reading the same receipts we have will answer both without looking anything up.

Where This Leaves Us

Both companies shipped excellent models in the same week and charged the same price. That should have been a boring story, and for most of Friday it was.

What I took from it is that a per-token price no longer tells you much. The bill depends on an effort setting the vendor chose for you, a cache you may or may not be hitting, and an input threshold most people do not know exists. Pick either model; they are close. Then go and find those three settings in your own config.

For a business buying software, the useful question is no longer which model your team uses. It is whether they know where those settings are, wrote down where they set them, and kept your project moving on a Thursday morning when the dials stopped responding.

Nobody puts that on a launch slide. It was the only thing on our own invoices this month that we could control.

At Rocky Soft, we build production web and mobile applications with Next.js, React, Node.js, NestJS, and React Native for clients in Calgary and across Canada. We use several frontier models on purpose and itemize what they cost. If you want software built by a team that reads the fine print on its own tools, let's talk about your project.

Frequently Asked Questions

Is GPT-6 Astra better than Claude Fable 5.1?

On the benchmarks both companies published, they trade wins. Astra leads on Terminal-Bench 4.0 (57.7% to 55.8%), Terminal-Bench Science, AutomationBench, and computer-use tests like ScreenSpot-Pro. Fable 5.1 leads on Humanity's Last Exam with tools (65.0% to 57.2%) and on the independent Artificial Analysis Intelligence Index, where it scored 66 to Astra's 61 on index version 4.1.1. On the Artificial Analysis Coding Agent Index, Fable 5.1 in Claude Code scored 70 against Astra in Codex at 67. Every's three-day hands-on comparison found Fable better at building simple, correct products and writing, and Astra better at being redirected mid-task. Neither company published a SWE-bench Verified score. For most software work the two are close enough that how you configure and route them matters more than which one you pick.

Which is better for coding, GPT-6 Astra or Claude Fable 5.1?

For terminal-heavy, browser-heavy, and long-running automation, GPT-6 Astra has the edge, and it tends to finish with fewer output tokens. For ambiguous product work where judgment and taste matter, Every's three days of side-by-side testing preferred Claude Fable 5.1. CodeRabbit's code-review evaluation put Astra at 61.3% actionable bug coverage against 50.2% for Claude Opus 5, but did not test Fable 5.1. Cognition, which runs the Devin coding agent, measured Astra within 0.4 points of Fable 5 on its benchmark at 64% lower cost. Both lead Opus 5 on most published rows, and Opus 5 still costs half as much.

How much does GPT-6 Astra cost compared to Claude Fable 5.1?

The list price is identical: $10 per million input tokens and $50 per million output tokens for both, with batch processing at half price. The differences are in the fine print. Fable 5.1 cache reads cost $0.25 per million against Astra's $1.00, which Anthropic says makes typical workloads about 25% cheaper and agentic workloads up to 45% cheaper. Astra charges a long-context premium: any request over 272,000 input tokens is billed entirely at $20 in and $75 out. Astra also offers a fast mode at double price. In practice, Artificial Analysis measured Astra at $1.67 per task and Fable 5.1 at $3.76 on its Intelligence Index at max effort, because Fable emits far more tokens. On other workloads the ordering flips. The effort level you choose changes the bill by up to 25x on Fable 5.1 and roughly fourfold on Astra.

Should I switch from Claude to GPT-6?

Probably not on the strength of one launch week, and not on benchmarks alone. Astra's rollout was uneven, with a launch post that went offline, enterprise security customers getting access before paying subscribers, weekly caps of 50 or 200 messages depending on which Pro plan you hold, and ChatGPT Plus users limited to Codex and the Work tab as of September 7. If your current stack works, our advice is to give any new model two weeks before it touches deadline work, then run it on your own tickets rather than trusting anyone else's numbers. If you are already multi-vendor, add Astra to the rotation for terminal and browser tasks. If you are single-vendor on either company, the more useful change is to stop being single-vendor.

When was GPT-6 released and is it available in Canada?

OpenAI released GPT-6 Astra in a limited rollout on Thursday, September 3, 2026 and broadened it on September 4. Claude Fable 5.1 was released two days earlier, on September 1. Both public models are available in Canada through their vendor APIs and Microsoft Foundry, Fable 5.1 is also on Amazon Bedrock and Google Cloud, Astra reached Bedrock on September 8, and there is no export restriction on either. What is restricted is the top tier: Claude Mythos 5.1, the same model as Fable 5.1 with fewer safeguards, is available only to a set of US organizations, and OpenAI's fuller cyber capability goes to vetted defenders through its Daybreak program. Both vendors sell regional-processing options at roughly a 10% premium, but for US or EU processing, not Canada.

Does the AI model my developer uses change what my software costs?

Less than you would think at the list-price level, and more than you would think everywhere else. Both flagships cost the same per token. What varies is how many tokens each one uses to finish a task, what effort level the agent runs at, whether prompts are structured to hit the cheaper cache, and whether anyone is watching for long-context surcharges and safety reroutes. The same prompt can cost 13 cents or $3.30 on the same model depending on one setting. A well-run shop treats those settings as engineering decisions and can show you what a piece of work cost and why. A poorly run one quotes a per-token rate and passes the surprise along.

Which AI model should my business use in 2026?

For most businesses buying software rather than building models, the answer is several. Claude Opus 5 at $5 in and $25 out remains the best value at the top for general development. GPT-6 Astra is the pick for browser automation, terminal-heavy agents, and long-context work under 272,000 tokens. Claude Fable 5.1 is the pick for ambiguous, judgment-heavy tasks where its extra tokens buy better first drafts. Gemini 3.8 Flash at $0.75 in and $3.75 out, and open-weight models like GLM-5.3-Flash or Qwen 3.8 27B on your own hardware, handle the high-volume routine work at a fraction of the cost. The more important decision is to route work deliberately and write the routing down, so a vendor outage or a pricing change is an adjustment rather than an emergency.

What happens to my project if one AI model gets switched off again?

It depends entirely on whether your development team planned for it. Claude Fable 5 was unavailable for nineteen days in June 2026 under a US government directive. On September 3, 2026, Claude was down for about three hours, and ChatGPT, Codex, and Grok were offline for part of the same window. OpenAI is cutting Cursor off from its models on November 12 over a corporate dispute. A team with a written routing policy, a second frontier vendor, and a local open-weight model for routine work keeps shipping through all three. A team built around one vendor's tool stops. Ask your developer when they last tested their fallback, and what the answer was.

Sources

• Anthropic. (2026, September 1). "Introducing Claude Fable 5.1 and Claude Mythos 5.1." Read announcement
• Anthropic. (2026). "Claude Fable 5.1 model overview." Read documentation
• Anthropic. (2026). "What's new in Claude Fable 5.1." Read migration notes
• OpenAI. (2026, September 3). "GPT-6 Astra: A new generation of intelligence." Read announcement
• OpenAI. (2026). "GPT-6 Astra model reference." Read documentation
• OpenAI. (2026). "API pricing." Read pricing
• OpenAI. (2026, September 3). "GPT-6 Astra System Card." Read system card
• Artificial Analysis. (2026, September 3). "Benchmarking GPT-6 Astra." Read analysis
• Artificial Analysis. (2026, September 1). "Claude Fable 5.1." Read analysis
• ARC Prize. (2026, September 3). "GPT-6 Astra on ARC-AGI-3." Read results
• Willison, S. (2026, September 1). "Claude Fable 5.1 made me a really nice animated pelican." Read post
• Willison, S. (2026, September 3). "GPT-6 Astra." Read post
• Parrott, K. (2026, September 6). "A split verdict on Fable vs. Astra." Every. Read review
• Thorelli, E. and Al-Hossami, E. (2026, September 4). "GPT-6 Astra review: code review gains, privacy, and cost." CodeRabbit. Read evaluation
• Mowshowitz, Z. (2026, September 8). "Astra Is Hard To Monitor." Read post
• Mowshowitz, Z. (2026, September 5). "Claude Mythos 5.1 and Fable 5.1: Capabilities." Read post
• Latent Space. (2026, September 4). "GPT-6 Astra: OpenAI's biggest LLM launch of all time." Read newsletter
• Swain, G. (2026, September 7). "Sam Altman calls GPT-6 Astra rollout 'messy' as enterprise users wait for access." CSO Online. Read article
• Zahn, S. (2026, September 7). "GPT-6 Astra is on ChatGPT Plus, but only in Work and Codex." Notebookcheck. Read article
• paddo.dev. (2026, September 5). "GPT-6 Astra: Critical, and generally available." Read post
• Doshi, T. and Popa, R. A. (2026, September 2). "Gemini 3.8 Flash and 3.8 Flash Cyber." Google. Read announcement
• Moon, M. (2026, August 29). "OpenAI will pull its models from Cursor due to the SpaceX acquisition." Engadget. Read article
• METR. (2026, August 26). "Investigation of the OpenAI Hugging Face incident." Read report
• Mazmanov, D. (2026, September 3). "Portal by Spotify cut my Claude Code token usage by 90%." Spotify Engineering. Read post
• Anthropic. (2026, September 3). "Elevated errors for multiple models." Claude Status. View incident
• NVIDIA. (2026, September 3). "NVIDIA to acquire Hugging Face." Read announcement
• Anthropic. (2026). "Claude Code May-August 2026 weekly limits promotion." Read help article
• Rehberger, J. (2026, August 26). "Breaking Claude Code Opus 5 Auto Mode." Embrace The Red. Read post
• Hacker News. (2026, September 3). "GPT-6 Astra." View thread
• Hacker News. (2026, September 1). "Claude Fable 5.1 and Claude Mythos 5.1." View thread