Quick answer: use Opus 5.5 by default and treat Fable 5.1 as the exception.
Anthropic says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work," and it costs 60% less per token. I switched my own default from Fable to Opus 5.5 today, and the logs are below.
Then I tested it on something my wife and I actually want. One of the biggest things we hate: when we play on the Switch, we're playing by ourselves. It kind of sucks.
So I asked Opus 5.5 for a Pokémon game we could play together. It didn't build one. It found one on GitHub, and that turned out to be the smarter move.
My rule now: Opus 5.5 by default, Fable only when Opus 5.5 at max effort still falls short.Free
Pro ($20/mo)
Max (from $100/mo)Opus (now Opus 5.5)
No
Yes, inside plan limits
Yes, inside plan limitsFable 5.1
No
Usage credits only
Up to 50% of weekly limitsSonnet and Haiku
Yes
Yes
YesPlan rows from Claude's pricing page and the Fable plan article, read September 22, 2026.
I switched my default to Opus 5.5 today, and the log shows the minute
I pay for Claude Max 20x. In Claude Code, every model change prints one line, and those lines stay in the session logs. Here are mine.My Claude Code /model log, September 17 and 22, 2026. Exact text, new frame.
Fable 5.1 at medium had been my saved default since September 17. This morning I dropped Fable to low, then tried Opus 5 at max effort. At 18:08 UTC, Opus 5.5 became the default for new sessions, at high effort.
That effort choice is on purpose. With Fable, I'll use medium. I always got good results. I use high, but very rarely.
Opus 5.5 is a step up from Fable 5.1 on paper, and the results people are claiming are actually cool, so I'm running it at high. I'm testing it myself. So far it's easier to talk to: when I showed it a hype post, it just called it hype.
The desktop app looks like this on my account today.My Claude desktop app, Max 20x, September 22, 2026.
Two things in that picture matter. Fable 5.1 is still the first name on the list, labelled "For your toughest challenges," and Opus 5.5 is "For complex tasks." And the app's own effort warning says Max effort means "4× or more usage," so the effort dial matters as much as the model name.
Opus 5 did not disappear either. Anthropic's deprecation page lists it as active until at least July 24, 2027, and the pricing page lists it under legacy models.
If you do not see Opus 5.5, update. In Claude Code, run claude update: Anthropic's model docs say Opus 5.5 needs version 2.1.280 or later. In the desktop app, update the app. On the web version of Claude, it does it for you.
Anthropic's Claude Code docs now list Opus 5.5 as the default model on Pro, Max, Team and Enterprise.
The game test: Opus 5.5 found a game instead of building one
Launch day comes with launch-day claims. Within an hour of the release, an X post said Opus 5.5 had one-shot a Mario Kart game.
I wanted to know how real that was, because I've wanted a game for me and my wife to play. So right after the switch, I handed Opus 5.5 the post and asked.My Claude Code session, September 22, 2026, model claude-opus-5-5. Short excerpts; [...] marks cuts.
It read the replies and called it straight. People asked for the prompt, the time and the tech stack, and got no answer, so treat "one shot" as marketing.
Then I asked for what I actually want, a Pokémon Yellow where you can catch everything. Instead of starting from zero, it found a fan remake that already does most of it, Pokémon Yellow Legacy on GitHub. It built it on my Mac and played the opening itself.
Did I build a game? No, we don't have to lie. It found one already on GitHub and gave me a suggestion to modify it. I have the logs to prove it.
It also checked the rest of my wishlist against what exists:Kanto and Johto in one game: Yellow cannot hold a second region, so that needs a Crystal base (it pointed at Polished Crystal).
Playing together online: a different game entirely, PokeMMO.
Phones: yes, through the Delta app on iPhone.
Switch: skip it.That list is how we landed on the game we're building right now. The Kanto-and-Johto wish pointed at a fan project on GitHub that rebuilds Kanto on Pokémon Crystal's engine, and we're building on top of it: a version my wife and I can play together.Pikachu following me in Viridian City, September 22, 2026.
Today Opus 5.5 added the first piece. There's a "who's playing" screen, a separate save for each of us, controller support, and the server that will put us in the same game at the same time. That last part is still being built.The picker Opus 5.5 added today, and Pallet Town. The Kanto-on-Crystal base is a fan project; two-player is ours.
That is the honest version of the one-shot claim. The move that worked on day one was find what already exists, then change it.
The other job was less fun and more useful. My Substack Notes run had been failing, and Opus 5.5 took it over today; the session log shows "10 notes are live in Notion." That's one run, so it's too early to call.
Pro just got Fable-level work inside the $20 plan
On Pro, Fable was never really inside your plan. Anthropic's support article says Fable on Pro and standard Team seats is not included in the plan's usage limits and runs on usage credits, which you pay for on top.
Opus is included on Pro. So the moment Opus 5.5 landed at Fable 5.1's level on most work, a $20 subscriber got that level of output without touching a credit balance.Claude's pricing page, September 22, 2026. Each cell says "Opus", not "Opus 5.5".
On Max, nothing about Fable changed today. It is still capped at up to 50% of your weekly limits and it draws from the same pool as everything else. Opus 5.5 carries no separate ceiling.
If you are still deciding which plan to be on, Claude Pro vs Max runs the same logic on the plan itself.
Price: 20% cheaper per token, 40% cheaper in Anthropic's tests
There are two different numbers here. Both come from Anthropic's launch page.Model
Input per 1M tokens
Output per 1M tokens
Opus 5.5 comparedOpus 5.5
$4
$20Opus 5
$5
$25
20% cheaper per tokenFable 5.1
$10
$50
60% cheaper per tokenThe 20% is the list price cut against Opus 5. The 40% is different: "Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads."
Against Fable 5.1 the gap is bigger. Fable costs 2.5 times what Opus 5.5 costs per token. Anthropic also says Opus 5.5 writes output more than 30% faster than Opus 5.
One warning for subscribers. Those are API prices. Anthropic has not published how much of your Pro or Max limit an Opus 5.5 message uses compared with a Fable message, so nobody can honestly tell you it stretches your week by a set amount.
Benchmarks: Opus 5.5 beats Fable 5.1, and Astra still wins two rows
On Anthropic's own benchmark table, Opus 5.5 beats Fable 5.1 on every row. A few that map to real work:Terminal-Bench 4.0 (agentic coding): 66.4% vs 55.8%.
GDPval-AA (knowledge work): 1846 vs 1735.
OSWorld 2.0 (using a computer): 81.8% vs 80.7%, basically level.It does not beat everything. OpenAI's GPT-6 Astra wins AutomationBench, 41.4% to 40.0%, and Terminal-Bench-Science, 64.6% to 58.7%. If Astra is the name you keep hearing, what GPT-6 Astra is covers it.
Anthropic even undercuts its own table: "In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest."
On Hacker News, sailingparrot did the math in the launch thread: "Fable 5.1 came out just 21 days ago. Only 3 weeks! And this is 20% relative improvement on terminal bench vs Fable 5.1 at less than half the price, and more human sounding output."
The writing point is the one I care about, because I use this for content. Anthropic says Opus 5.5 "communicates more naturally than prior models."
mcintyre1994 put the frustration plainly in the same thread: "I mostly moved to Astra because I just can't work all day with the Claude Opus 5/Fable writing style... Definitely keen to try Opus 5.5 and see if this claim is real."
Fable still wins one kind of job, and it is rarer now
Here is where Fable keeps its seat. Anthropic's model guide for developers says "Most workloads start with Claude Opus 5.5." Then it gives one exit: move to Fable 5.1 if your results "at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work."
In plain English, Fable is for the long job that runs for hours and has to hold a lot together. Anthropic's examples are agent sessions that run for hours and multistep deep research.
Before today, that was my rule too. Fable plans it, Opus makes it is how I ran things last week.
Today my answer changed. I wondered out loud whether I should keep using my Fable usage because it was separate. It is not separate. On Max it comes out of the same weekly limits.
So why keep paying more? Do you want to pay more money for something less? Doesn't make sense, right? If Opus 5.5 is better all around, it makes literally no sense to use Fable as of now.
I don't think that lasts. When Anthropic ships a Fable 5.5, I expect it to be even more powerful, and the math flips back.
Your five-hour limit went up, and nobody has a number
Anthropic's launch page says it is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans." It published no figure, so any percentage you see online is a guess.
There is also a new rate limit reset for subscribers, "which you can now save and use whenever you choose." Think of it as a spare refill you get to time yourself.
How the five-hour and weekly windows actually work is on Claude usage limits explained.
Which one to pick, by planIf you are on Free: neither. Free is Sonnet and Haiku, and Opus 5.5 is not on it.
If you are on Pro at $20: Opus 5.5 for everything. Fable would cost you credits for work Anthropic says is level on most jobs.
If you are on Max: Opus 5.5 by default. Save Fable for the rare job that runs for hours, and only after Opus 5.5 at max effort fell short.
If you write or create for a living: Opus 5.5. Anthropic says it communicates more naturally than prior models, and it is the model I moved my content work to today.
If you were running Opus 5 at max effort to squeeze out quality: try Opus 5.5 at high first. That is the swap my own log shows.
If the plan question is still open: Claude Pro vs Max is the next read, and the Claude Guides hub maps the rest.Common questions
Is Claude Opus 5.5 better than Fable 5.1?
On most work, Anthropic says they are level, and Opus 5.5 beats Fable 5.1 on every row of Anthropic's benchmark table, for example Terminal-Bench 4.0 at 66.4% vs 55.8%.
Anthropic still tells developers to move to Fable 5.1 when Opus 5.5 at xhigh or max effort falls short on demanding reasoning or long-horizon agentic work.
Does Claude Pro get Opus 5.5?
Yes. Claude's pricing page marks Opus as included on Pro, Max 5x and Max 20x, and not on Free. Fable on Pro is marked Usage credits, billed on top of the plan, so Opus 5.5 is the strongest model inside Pro's limits.
Is Opus 5.5 cheaper than Fable 5.1?
Per token, yes. Opus 5.5 is $4 input and $20 output per million tokens and Fable 5.1 is $10 and $50, so Opus 5.5 is 60% cheaper per token. Anthropic has not published how much of a subscription limit either model uses per message.
Is Fable usage separate from my Opus usage on Claude Max?
No. On Max, Fable draws from the same weekly usage limits as every other model and is capped at up to 50% of them. On Pro and standard Team seats, Fable is outside the plan's limits and runs on paid usage credits.
Why can't I see Opus 5.5?
Update. In Claude Code, Opus 5.5 needs version 2.1.280 or later, so run claude update. In the desktop app, update the app; the web version updates by itself. Opus 5.5 is not on the Free plan.
Did Anthropic retire Opus 5?
No. Anthropic's deprecation page lists Opus 5 as active until at least July 24, 2027, and the pricing page lists it under legacy models. Picking it is still possible, it is just no longer the top Opus.
Where this comes from
The facts come from Anthropic's own pages, read on launch day, September 22, 2026. The logs, the picker and the Pokémon session are mine, from my Max 20x account that same day.
I have not run my own benchmark or measured my limits since the raise. It's day one, so I'll update this as I go.
Published September 22, 2026, launch day. Last reviewed September 22, 2026.
Quick answer: use Opus. As of September 22, 2026, Anthropic says Claude Opus 5.5 "performs at the level of Claude Fable 5.1 on most work", and it costs 60% less per token.
It sits inside your Pro and Max limits. Fable still runs on paid usage credits on Pro and is capped at 50% of your weekly limits on Max.
I'm used to switching models between Fable and Opus, and Fable was always the upgrade. Now it's the opposite: Fable is the exception you have to justify.
This page was updated September 22 for that launch. The launch itself is covered in Opus 5.5 vs Fable.
Anthropic's own numbers and your plan now point in the same direction.
My rule on the Claude picker, since September 22: we're gonna use Opus now.
The picker has too many names, and Opus 5.5 is the one to leave ticked
Open the model picker in the Claude desktop app and the top of the list reads Fable 5.1, Opus 5.5, Sonnet 5 and Haiku 4.5. Under them sit an Effort control and a More models menu.
Then ChatGPT starts shouting Sol and Astra at you from the next tab over, and if you want that particular headache sorted, Astra vs Fable does it.
When you have ten models to pick from, it is just too much. What the hell is Sol. What is Astra.That is my own picker on September 22, 2026, with Opus 5.5 ticked and Effort on Medium. Notice Anthropic's own labels still pitch Fable 5.1 "For your toughest challenges" and Opus 5.5 "For complex tasks".
The label has not caught up with the launch page. Opus 5 has not gone anywhere either: Anthropic's pricing page now files it under legacy models.
Effort is how hard Claude thinks before it answers: raise it and you get more reasoning and spend more of your limit, lower it and you get a faster, cheaper answer. It matters as much as the model name, and it shows up again at the end of this page.
So here is the compression. Until September 22 I told you to learn two names: Opus for the work, Fable for the thinking. Now it is Opus 5.5 by default, and Fable only when you can say why. Everything else is detail.
I did the four names in a minute on camera, before Fable 5.1 landed:The 60-second version, from my Shorts: Claude's 4 models in plain English (and the one you're wasting).The answer in one table: what each plan actually gets
Read this before you click anything. The model rows come from Anthropic's pricing page and the Max cell's 50% wording comes from their Fable support article. Both were first read September 18, 2026 and checked again September 22, after the Opus 5.5 launch.Free
Pro ($20/mo, $17 annual)
Max (from $100/mo)Sonnet
Yes
Yes
YesHaiku
Yes
Yes
YesOpus
No
Yes
YesFable
No
Usage credits
50% of weekly limitsUsage credits available
No
Yes
YesTwo cells in that table are the whole page. Fable on Pro reads "Usage credits", not a checkmark. Fable on Max reads "50% of weekly limits", not "unlimited".
None of the rows moved on September 22. Opus 5.5 landed in the same Opus row, so Pro and Max get it inside their limits, and the Fable cells are exactly where they were.
Anthropic also said it is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans", with no numbers, plus a rate limit reset "which you can now save and use whenever you choose."
On Pro, Fable is a paid add-on wearing an included label
Here is the sentence that every "Fable is on all paid plans" summary stops one line short of.Anthropic's headline is accurate: "Claude Fable 5 and Claude Fable 5.1 are available on all paid plans (Pro, Max, Team, and Enterprise)." If Fable itself is new to you, what Claude Fable is covers the model rather than the billing.
Then the detail. For Pro plans and standard Team seats: "Fable 5 and Fable 5.1 aren't included in your plan's usage limits. You can use them with usage credits, which let you pay for usage beyond what your plan includes."
Available is not the same as included.
If you are a $20 subscriber clicking Fable expecting it to come out of the allowance you already pay for, you are about to meet a paywall, and page one of Google does not warn you.
There was a one-time credit for eligible Pro and Team standard seats when Fable 5 moved to usage credits in July 2026, and Anthropic is explicit that there is no equivalent for Fable 5.1.
There was also a promotion that gave subscribers up to 50% of their weekly subscription limit on Fable 5 at no extra cost.
Per the same support article, it ended on July 19, 2026, and it never applied to Fable 5.1. If you read a guide saying Pro gets 50% of its limits on Fable, you are reading something written before that date.
On Max, the 50% is a ceiling and not a bonus
The Max rule gets restated online as a perk. It is a cap.
Anthropic: "You can use up to 50% of your weekly usage limits on Fable models at no extra cost. They draw from your plan's regular weekly usage limits and use them faster than other Claude models."
Fable does not come with its own tank. It drinks from yours, faster than anything else does, and once half the week is gone that is the end of the free part. After that, Anthropic says you can keep going on usage credits or switch models.That is my own account on September 18. Max 20x, and a usage credits balance sitting there as the paid-on-top lane for exactly this situation. Card details are blurred, the mechanism is not.
Here is what the ceiling feels like in practice. One day into my week I was already at 20%, and Fable is most of why. I use Opus heavily as well, and I do not use anything below it.That meter capture is from September 5, 2026, not today. It is what a real Max week looks like when Fable is in the rotation. The full mechanics live on Claude usage limits explained.
Nobody publishes how much faster Fable burns, including Anthropic
I want to be blunt about this because the internet is full of confident numbers.
Anthropic says Fable uses your limits "faster than other Claude models" and, in their model guide, that it "uses the most of your rate limit". Their ladder is: Haiku lightest, Sonnet moderate, Opus heavy, Fable heaviest.That table is as of September 18, so it still names Opus 5, and the guide had not been updated for Opus 5.5 when I checked on September 22. That is the whole published guidance. No multiplier. No "Fable is 5x Opus".
Anthropic has not said where Opus 5.5 sits on that ladder either. The token price is lower, but how that shows up on your meter is not published, so I am not going to tell you Opus 5.5 uses less of your week than Fable.
So my "a day in and I am at 20%" is a dated personal observation on Max 20x with my particular workload. It is not a rate, and you should not plan your month around it.
Even Claude's own tooling cannot keep this straight
If you find this confusing, you are in good company.
On July 20, 2026 a Max subscriber going by tobowers filed issue #79412 on the Claude Code repository with a title that says everything: "/model shows Fable 5 as credits-only (€0.00) on Max plan, while /usage shows Fable covered by plan (5% of weekly limit)."
In the same session, the picker told him "Fable 5 runs on usage credits, you have €0.00 in credits" while /usage reported "Current week (Fable): 5% used."
His own read: "The Current week (Fable) bucket is a plan-usage allowance that resets on the weekly cycle, i.e. Fable is included in my plan and I'm 5% into it."
He was right, and the tool was telling him two opposite things at once. The issue is marked as a duplicate and I can see no Anthropic response on it, so I am not going to claim it was confirmed or resolved.
My rule now: use Opus, and make Fable earn it
This is what I get asked about most, and it flipped on September 22.
The new rule is one line. We're gonna use Opus now.
If Opus 5.5 is better all around, it makes literally no sense to use Fable as of now.
Do you want to pay more money for something less? It doesn't make sense, right?
My Claude Code log has the switch: Opus 5.5 saved as my default for new sessions on September 22. That is not me retiring Fable. It is Fable losing the default spot.
I even caught myself thinking my Fable allowance on Max was a separate bucket. It isn't. It is a cap of up to 50% inside the same weekly pool as everything else.
Anthropic's developer guide says the same thing in drier words: "Most workloads start with Claude Opus 5.5." It only sends you to Fable 5.1 if Opus 5.5 at its top effort settings (xhigh or max) still falls short on demanding reasoning or long-horizon agentic work.
That guide is written for people building on the API, but the order translates. Turn the effort up on Opus 5.5 before you reach for Fable.
Anthropic still describes Fable as being for "your largest, most critical projects: long, complex tasks Claude works through with fewer check-ins", and it takes longer to answer. Next to the developer guide, that makes the exception narrow: a long, demanding job where Opus 5.5 at its top effort already fell short.
And I know it will flip again. Fable's gonna become the big dog eventually, when they release Fable 5.5 or whatever they call it. I bet they have a stronger model out there already, testing and testing. That is my prediction, not anything Anthropic has announced.
Until September 22, my rule was: Fable plans it, Opus makes it
For four days this page said the opposite, and I meant it. On September 18 my rule was Fable for planning and orchestrating, Opus for producing the thing once the engine is built.
In my words that day: "So I use it a lot for planning and thinking, big stuff." And for content: "You're not gonna use Fable to create a tweet, you're just gonna use Opus."
For planning work I would rather have run Fable at medium effort than push Opus 5 to high or max. That was a preference, not a measurement, and it was about Opus 5. Opus 5.5 is a different model, and Anthropic puts it at Fable 5.1's level on most work.
One part of the old rule still stands. I do not trust my own judgment in the moment, so I put the routing in a rules file so it happens every time instead of whenever I remember. If you find yourself deciding model-by-model, you will default to the expensive one when you are tired.
Why I do not just use Sonnet
Every page like this gets a comment saying just use Sonnet. Here is the honest answer to that.
The thing that degrades is not the paragraph. It is the number of round trips.
Do research on a cheaper model and you get high-level research. Which means you go back. "This is not what I asked for." "You are going in the wrong direction." And you churn.
That churn is invisible in a price comparison and expensive in a real week. When I do not want to waste time, and I want a partner smart enough to actually think with, I reach past Sonnet.
Until September 22 that meant Fable. Now it means Opus 5.5. Fable earns the burn only when Opus 5.5 at its top effort still falls short on long, demanding work, which is Anthropic's own escalation rule.
Which is also the honest counter-argument to my own rule: if your work does not involve holding a complicated system in your head, you will never feel this, and Sonnet will be fine.
Anthropic's consumer model guide still recommends Sonnet as the default for anyone unsure, and it is the only strong model on the free plan.
The other reason page one is not helping you
Every comparison ranking for this question prices it in API tokens. Second Talent's model roundup is a good example, and a fair one: "Reach for Fable 5 when the task is the hardest you have and the cost of a wrong answer is high", alongside a price of "$10 input and $50 output per million tokens".
That was a fair answer to a different question, and it was written about Fable 5, before Opus 5.5 existed. Here are the list prices from Anthropic's pricing page on September 22, 2026:Per million tokens
Input
OutputOpus 5.5
$4
$20Fable 5.1
$10
$50Opus 5 (now listed as legacy)
$5
$25Fable 5.1 used to be twice the price of Opus 5. Against Opus 5.5 it is 2.5 times the price.
You do not buy tokens, though. On Max you have a subscription and a percentage that goes down; on Pro you have a subscription plus a credit balance that Fable draws against.
Either way, "2.5 times the price per million tokens" means nothing when your real currency is how much of Thursday you have left.
So which one should you pickIf you are on Free: neither. You have Sonnet and Haiku, and Sonnet is good. See the Claude free plan.
If you are on Pro at $20: use Opus 5.5 and stop there. It is inside your plan, and Anthropic puts it at Fable 5.1's level on most work. Fable is a paid extra, so check your credits before you click it, not after.
If you are on Max: Opus 5.5 by default. Fable only for a job where Opus 5.5 at its top effort already fell short, and watch the weekly meter, because half your week is the cap.
If you write or create for a living: Opus. Fable costs 2.5 times as much per token, runs slower, and will not make your sentences better.
If you are designing a system, not using one: until September 22 this was the case for Fable at medium effort. Now start on Opus 5.5 and turn the effort up first. Fable is the fallback if that still falls short.
If you genuinely cannot decide: Opus 5.5 if you pay. Anthropic's consumer guide still says Sonnet, and Opus vs Sonnet is that fork if you are on the fence.Two other decisions sit next to this one. If the question is really which plan to be on, Claude Pro vs Max answers it with the same weekly-allowance logic.
And the Claude Guides hub is the map for everything downstream.
Common questions
Is Fable included in Claude Pro?
Not inside your plan's usage limits. Anthropic's support article says Fable 5 and 5.1 are not included in Pro's usage limits and that you use them with usage credits, which let you pay for usage beyond what your plan includes. Anthropic's pricing table marks Fable on Pro as "Usage credits".
How much Fable do I get on Max?
Up to 50% of your weekly usage limits at no extra cost. That is a ceiling inside your existing allowance, not additional capacity, and Fable draws from that pool faster than other Claude models.
What happens when I hit the Fable ceiling?
You can keep using Fable on usage credits, or switch to another Claude model to stay inside your plan's limits. Those are the two options Anthropic lists.
How much faster does Fable burn than Opus?
Anthropic does not publish a figure. They rank rate-limit use as Haiku lightest, Sonnet moderate, Opus heavy and Fable heaviest, and say Fable uses the most. Any specific multiplier online is an estimate.
Should I use Fable or Opus for writing?
Opus. As of September 22, 2026, Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, and it costs 60% less per token. Fable is slower and eats your weekly allowance faster, and a post or a tweet does not need it.
Which Claude model should I pick if I do not want to think about it?
Opus 5.5 if you pay for Claude. Anthropic's developer guide now starts most workloads on Opus 5.5. Its consumer model guide still names Sonnet as the default and had not been updated for Opus 5.5 as of September 22, 2026. Save Fable for jobs where Opus 5.5 at its top effort still falls short.
Is Fable better than Opus?
Not on most work, as of September 22, 2026. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work and beats it on every row of its launch table.
Fable is now the fallback when Opus 5.5 at top effort falls short on long, demanding jobs. It costs 2.5 times as much per token.
Is Opus 5.5 better than Fable?
On most work it is at least level, per Anthropic. Opus 5.5 scores higher than Fable 5.1 on every row of Anthropic's launch table, such as 66.4 vs 55.8 on Terminal-Bench 4.0.
It costs $4 and $20 per million tokens against $10 and $50. That is most work, not every job, and Anthropic publishes no usage-limit comparison.
Where this comes from
The plan rules are Anthropic's pricing page and Fable support article, read September 18 and again September 22. The Opus 5.5 claims are Anthropic's launch page and developer guide, read the day it shipped.
The account screenshots are my own Max 20x plan, dated. I haven't measured Opus 5.5's burn against Fable, and Anthropic publishes no number.
Correction log: what this page has said, and whenSeptember 18, 2026. Published with the rule "Opus for the work, Fable for the thinking": Fable at medium effort for planning and system design, Opus for producing the work.
September 22, 2026. Anthropic released Opus 5.5 and said it performs at the level of Fable 5.1 on most work. Replaced the rule with "use Opus, and make Fable earn it" and kept the old one above as dated history. Swapped the September 5 picker screenshot for a September 22 one, re-dated the plan table (no rows changed), moved the per-token gap from 2x to 2.5x, rewrote the writing, default and "Is Fable better than Opus?" answers, and added "Is Opus 5.5 better than Fable?"Published September 18, 2026. Updated September 22, 2026 for Opus 5.5, with the dates in the correction log above.
Quick answer: these two models now cost exactly the same and are good at different things. GPT-6 Astra wins computer use, math, cybersecurity and long context. Claude Fable 5.1 wins the two broadest "how smart is it" boards, including one published by OpenAI itself.
Both are $10 per million input tokens and $50 per million output. So price isn't what decides it anymore.
Every headline yesterday said Astra swept. OpenAI's own launch table does not say that.
The rows nobody quoted
OpenAI published a comparison table on the Astra launch page with Claude Fable 5.1 in a column. Two lines in that table go to Claude.
Every number below is transcribed from that page, checked September 4, 2026. When OpenAI's footnotes put a caveat on a Claude score, I flag it in the next section.Benchmark (OpenAI's own table)
GPT-6 Astra
Claude Fable 5.1Artificial Analysis Intelligence Index v4.1.1
61.2
65.7Humanity's Last Exam (with tools)
57.2%
65.0%Terminal-Bench 4.0
57.9%
55.8%Terminal-Bench Science 0.1
64.6%
52.6%FrontierMath Tier 4 (v2)
97.6%
87.8%GPQA Diamond
96.0%
93.7%AutomationBench
41.4%
31.4%DeepSWE v1.1
74.1%
67.4%FrontierCode 1.1 Main
53.3%
50.9%ARC-AGI-2
95.0%
90.0%Computer use safety, internal (lower is better)
2.4%
9.5%The Artificial Analysis Intelligence Index is about as close as you get to one number for how smart a model is overall. OpenAI put it in its own launch table, and it hands Claude a 4.5 point lead.
Artificial Analysis, quoted independently:Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback).So on the broadest board, OpenAI's newest model tied its last one.
Read the footnotes before you read the table
OpenAI ran these numbers. That does not make them wrong, but the footnotes change how you should read four rows.BenchCAD: OpenAI's footnote 5 says Claude's scores reflect three modifications to the eval.
ScreenSpot-Pro and ExploitGym: footnote 17 says the Fable scores reported are actually from Mythos, described as "Fable with fewer safeguards." That is a different model.
HealthBench Professional: footnote 11 says OpenAI independently evaluated the Claude models itself, using GPT-5.4 as grader, with Opus 5 substituted when Fable 5.1 refused.
Three science evals: footnote 12 says Claude Fable 5 and 5.1 are excluded from LifeSciBench, GeneBench Pro and MedChemBench because they refuse the majority of questions.None of that is shady. Every vendor benchmarks its own launch. But "state of the art across the board" is how the press summed it up, and the table doesn't say that.
The 99.9% asterisk
The number that traveled furthest was Astra saturating ARC-AGI-3 at 99.9%. It should count the least of anything in the launch.Two problems, both disclosed by the people who ran it:Harness. Per the ARC Prize blog, the 99.9% came from OpenAI's custom "Provider Adapter harness" at about $19K. The default ARC-AGI harness scored 62.7% at about $26K. OpenAI's own footnote 1 confirms it ran a modified responses API harness. That is a 37 point spread depending on plumbing.
No opponent. Claude Fable has no published ARC-AGI-3 result. So the biggest number of the launch has nothing to compare against.The Provider Adapter, in ARC Prize's words, "preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work." That's interesting engineering. But it doesn't tell you how smart the raw model is.
Simon Willison, writing the day it shipped, summed up the launch in one sentence: Astra "appears to score higher than Fable on most of OpenAI's self-reported benchmarks."
The key word there is self-reported.
Where Astra genuinely pulls ahead
Strip out the noise and Astra's wins are real. They just bunch up in a few areas.
Computer use. This is the big new thing in the launch. Astra scores 59.3% on Agents' Last Exam against 55.5% for Claude Opus 5, and OpenAI reports it hitting 72.6% on OSWorld 2.0 in roughly 47% less time per task than GPT-5.6 Sol.
Math and science. FrontierMath Tier 4 at 97.6% against 87.8%. Terminal-Bench Science at 64.6% against 52.6%. These are not close.
Cybersecurity. 100% on ExploitBench, 88.0% single-attempt on SRE-Bench binary reverse engineering. OpenAI says this crosses the Critical threshold in its own Preparedness Framework, so it put limits on the model at launch.
Long context. 100% on OpenAI's eight-needle test at 256K to 512K tokens, and 96.3% at 512K to 1M.
Token efficiency. This one matters more than the scores. On Agents' Last Exam, OpenAI reports Astra using about 65% fewer output tokens than Opus 5 at the highest-scoring settings.
Where Claude Fable 5.1 holds
The broad boards. Intelligence Index and Humanity's Last Exam, both from OpenAI's own table.
Coding, closer than the headline. Astra edges Fable 5.1 on FrontierCode 1.1 Main, 53.3% to 50.9%. But look one column over in OpenAI's table: Fable 5 scores 53.5%, ahead of Astra. On the Artificial Analysis Coding Agent Index, Astra posts 67.0 against 67.2 for Fable 5 and 68.1 for Opus 5. Coding is a wash.
Cost per task at equal quality. Artificial Analysis again: "Per task, the model is less than half the cost of Claude Fable 5, for the same score." That's where Astra's efficiency actually pays off. And it matters more than the trophy rows.
What I actually threw at it on day one
I pay for both stacks, and Astra is one day old as I write this.I didn't run a benchmark. I just ran my actual work, because that's the only test I trust.
The one that changed my mind about Astra was a prospecting task. I asked it to find companies in a specific niche worth reaching out to, and it went and searched forums and threads on its own without me telling it where to look.
What came back was about a month old. That kind of search usually pulls up threads from many months back.
For me that's the real upgrade: it did the digging on its own.Another Astra run from the same week, on my Pro account. I asked it to plan two weeks of paid-work experiments around my real numbers; it pulled live listings on its own and priced the options. The top of the answer has my personal figures in it, so it is trimmed. The chip in the composer is the part worth noticing: on this surface it is labeled GPT-6 Astra Ultra, not GPT-6 Pro.
Then I did the thing I would actually recommend: I ran the output back through Claude Fable to QA it, Fable rewrote the prompt, and I sent it back at Astra. The second pass looked good enough that it's now how I'd run this by default.
Now the other side. I've used Fable more, and for building it still feels more familiar and more dependable to me. But Fable 5.1 has also done some dumb things this week, mostly losing track of tools it definitely had access to. It's minor, but it's real.
My read on the scoreboard: I assumed it said Astra beat Fable outright. It doesn't, and that actually matches what I saw hands-on. Fable still feels better at the deep work. Astra feels newer at getting things done by itself.
The decision rule I would give someone today
Don't switch providers over this. Send each job to the one that's better at it.The job
OpenDeep building, writing in your voice, long careful work
Claude Fable 5.1Anything the AI should do on its own across apps and the web
GPT-6 AstraResearch and volume where you need lots of runs
GPT-6 AstraReviewing and QAing another model's output
Claude Fable 5.1Math, science, anything with a checkable right answer
GPT-6 AstraYou can only pay for one, and you are a generalist
ChatGPT (usage and resets decide it, not the score)I keep coming back to this: smartest isn't the flex anymore. A model that burns tokens to win a board is worth less to me than a leaner one that finishes the task. Both companies are getting pushed that way, and this launch is the clearest sign of it so far.
Verdict
Same price, split by job, and that split is the useful part. Astra is the better agent. Fable is still the better thinker on the broadest boards OpenAI itself published, and it's my daily driver for building.
If you were about to cancel Claude because of a headline, read the table first.
Working out which subscription this actually affects? Which ChatGPT plans get Astra, Claude Max vs ChatGPT Pro, and Claude vs ChatGPT for the everyday call. More of how I run both at Claude Guides.
Published September 4, 2026, one day after GPT-6 Astra shipped. Every benchmark figure was transcribed that day from OpenAI's launch page, and the ARC-AGI-3 harness figures from Simon Willison's write-up citing the ARC Prize blog, both linked above. These are vendor-run numbers on a vendor page and OpenAI's footnotes qualify several Claude scores; read the footnotes section before quoting any row. Benchmarks and prices move fast and this page will need re-checking.
Chris Alarcon 
