- Claude Fable 5.1 is the most capable overall. It leads both the intelligence and the coding indexes. If quality is what you are buying, buy this.
- GPT-6 Astra is the one that operates your computer. Nothing else in this comparison drives Blender, Final Cut or a browser the way it does. That is its whole case.
- Gemini 3.8 Flash is roughly a thirteenth of the price and three times the speed. For high-volume work where "good" beats "best", it is not close.
- Price per token is the wrong number. A model that costs a thirteenth as much but needs three attempts is not cheaper. Work out cost per finished task.
- All three shipped within days of each other in September 2026, so treat every number here as a snapshot with a date on it, not a permanent ranking.
In the first week of September 2026, three frontier models arrived almost on top of each other: OpenAI's GPT-6 Astra, OpenAI’s computer-operating model on 3 September, Anthropic's Claude Fable 5.1, and Google's Gemini 3.8 Flash.
If you are choosing between them, you have probably already found a dozen comparison tables. They mostly agree, and they mostly stop at the same place: a ranked list of scores.
A ranked list is not a decision. The model that wins a benchmark and the model you should pay for are frequently not the same one, because benchmarks measure capability and your invoice measures capability divided by how many attempts it took.
This page gives you the scores, then goes past them.
The short answer, by situation
If you read nothing else, read this table.
| If you are… | Use | Because |
|---|---|---|
| Writing or reviewing production code | Claude Fable 5.1 | Leads the coding index; the quality gap costs less than the debugging it saves |
| Automating work across desktop apps | GPT-6 Astra | The only one of the three built to operate software rather than describe it |
| Running something at high volume | Gemini 3.8 Flash | Roughly 13× cheaper and around 3× faster; the price gap swamps the quality gap |
| Doing frontier maths or science | GPT-6 Astra | Its strongest measured area |
| Doing security or red-team work | GPT-6 Astra | First model OpenAI classified Critical for cyber capability |
| Not sure, general knowledge work | Claude Fable 5.1 | Highest general intelligence score; safest default |
| On a tight budget, any use | Gemini 3.8 Flash | Start here and only move up if quality actually blocks you |
The rest of this page shows the working.
The numbers, side by side
These are third-party measurements from Artificial Analysis, not vendor claims. I have labelled them Reported for that reason — they were measured by someone independent, but not by me, and benchmark numbers move as harnesses change.
| Measure | GPT-6 Astra | Claude Fable 5.1 | Gemini 3.8 Flash |
|---|---|---|---|
| Intelligence Index | 61 | 66 | 59 |
| Coding Agent Index | 67.0 | 70.4 | 61.1 |
| Input price / 1M tokens | $10 | $10 | $0.75 |
| Output price / 1M tokens | $50 | $50 | $3.75 |
| Speed (tokens/sec) | ~87 | ~67–69 | ~305 |
| Computer use (OSWorld 2.0) | 72.6% | — | — |
Three things jump out of that table, and only one of them is about intelligence.
1. The quality gap is small. 66 against 61 against 59 is a real ordering, but it is not a chasm. On most everyday work you will not feel five index points.
2. The price gap is enormous. $50 against $3.75 per million output tokens is not a discount, it is a different category of spending. That is roughly 13×.
3. The speed gap is bigger than the quality gap. ~305 tokens per second against ~67 is the difference between a tool that feels instant and one you wait for.
Gemini 3.8 Flash's pricing is promotional through 31 December 2026. Budget on the assumption it changes.
The number nobody publishes: cost per finished task
Here is the mistake almost every comparison makes, including the ones ranking above this page today.
They compare price per million tokens. You do not buy tokens. You buy finished work.
"Gemini 3.8 Flash is 13× cheaper" is true and can still be the wrong conclusion.
If the cheaper model needs three attempts where the expensive one needs one, your real saving is not 13×. It is closer to 4×. Still good — but now you have to weigh it against three times the waiting, three times the reviewing, and the risk that attempt three is still wrong.
And if a task fails in a way you do not catch, the cost is not the tokens at all. It is the mistake.
The honest formula is simple:
(price per run × average attempts to get it right) + the cost of your time reviewing it
The second term is the one people forget, and for most professional work it is larger than the first.
Work that through and the picture inverts depending on volume:
- Low volume, high stakes — a contract review, a production migration, a client deliverable. Attempts are expensive because your time reviews each one. Buy the best model. The token cost is rounding error against an hour of your day.
- High volume, low stakes — classifying support tickets, tagging products, first-pass summaries. Attempts are cheap because a machine checks them. Buy the cheap fast model, and spend the savings on a verification step.
- The middle — this is where it is genuinely hard, and where you should actually run the test rather than read a table. Take twenty real tasks from your own work, run them on two models, and count attempts. That measurement takes an afternoon and beats every comparison article, including this one.
I have written the full pricing mechanics — the surcharge cliffs, caching, rate limits — in the GPT-6 Astra API guide.
GPT-6 Astra: the one that uses the computer
Astra's case is not that it is the smartest. It is that it does things the other two do not do at all.
It operates software directly. Not "writes a script for Blender" — opens Blender, moves through menus, drags objects. On OSWorld 2.0, the benchmark for operating a computer, OpenAI reports 72.6%.
That number is worth sitting with in both directions. It means roughly three in four computer tasks complete. It also means roughly one in four still fails, which is a very different product from one you can leave alone.
What it is genuinely best at:
- Cross-application work. The handoff between two programs — model in Blender, import into Unreal — is where ordinary automation breaks and where Astra's demos consistently land.
- Frontier maths and science. Its strongest measured domain.
- Security work. It is the first model OpenAI classified Critical for cyber capability, meaning it can find unknown vulnerabilities and build working exploits against defended systems largely by itself. That is a capability and a liability in the same sentence.
What to weigh against it: the launch rollout was, in Sam Altman's own word, "messy" — paying subscribers were locked out for days while access went to enterprise partners first. And a model that operates your desktop needs to be set up like infrastructure, not like a chatbot.
If you want to see what people are actually producing with it, I keep a running index of builds made with GPT-6 Astra, each credited to its creator.
Claude Fable 5.1: the most capable, plainly
If you want the best output and the bill is secondary, this is the answer.
Fable 5.1 leads both indexes in this comparison — 66 on intelligence against Astra's 61, and 70.4 on the coding agent index against Astra's 67.0. It is the same list price as Astra, at $10 in and $50 out per million tokens.
Where that lead actually shows up:
- Code you intend to ship. The coding gap is the largest quality gap in the table, and it compounds — a model that gets it right first time saves review, not just tokens.
- Long, careful reasoning. General knowledge work where being subtly wrong is expensive.
- Writing that has to hold an argument together across thousands of words without drifting.
Where it does not win: it is the slowest of the three at roughly 67–69 tokens per second, and it does not drive your desktop. If your problem is throughput or operating software, its quality lead is irrelevant to you.
Gemini 3.8 Flash: the one the maths keeps choosing
Gemini 3.8 Flash scores lowest of the three and is still the right answer surprisingly often.
It sits at 59 on intelligence and 61.1 on coding — genuinely behind, but not in a different league. Then the other two columns arrive: $0.75 in and $3.75 out per million tokens, and about 305 tokens per second.
Put plainly: it is roughly a thirteenth of the price and around four times faster than Fable 5.1, for maybe a ten per cent quality step down on these indexes.
That ratio makes it the default for:
- Anything you run thousands of times. Classification, extraction, tagging, routing, first-pass drafts.
- Anything a user waits for. Speed is a feature. A fast adequate answer often beats a slow excellent one in a product.
- Anything you are still prototyping. Cheap iteration means more iterations.
The catch to diary: that pricing is promotional through 31 December 2026. Build a business case on it and check the renewal date before you scale.
How to choose, by what you actually do
Rules of thumb, in the order I would apply them.
1. Does the task require operating software? If yes, it is Astra. The other two do not compete here, so the rest of the comparison does not apply.
2. Will a human read every output? If yes, quality beats price — go Fable 5.1. Your review time costs more than the tokens.
3. Will a machine consume most outputs? If yes, price and speed beat quality — go Gemini 3.8 Flash and add a validation step.
4. Still unsure? Start on Gemini 3.8 Flash. Move up only when you can name the specific failure that forced you to. "It might be better" is not a reason to pay 13×.
One more that applies to everyone: do not standardise on one model. The prices and the leaderboard both moved within a single week this month. Write your code so the model is a configuration value, not an assumption, and switching costs you an afternoon instead of a quarter.
What these scores do not tell you
Four honest caveats, because a comparison that pretends to certainty is worth less than one that does not.
- Benchmarks are harness-dependent. The same model scores differently depending on how the test is run — scaffolding, retries, prompt format. I have unpacked this at length in why the benchmark records carry asterisks. Two comparison articles quoting different numbers are often both right.
- These are third-party measurements. Artificial Analysis is independent, which is exactly why I trust it more than a vendor chart — but it is one methodology, and it is not your workload.
- A benchmark is not your job. The only test that settles this is twenty of your own real tasks, run on two models, with attempts counted. Nothing on this page substitutes for that afternoon.
- This is dated, deliberately. Every model here shipped in the first week of September 2026. A comparison of frontier models has a shelf life measured in weeks, and any page claiming otherwise is selling you something.
FAQ
Which is the best AI model right now?
On measured capability, Claude Fable 5.1 — it leads both the Artificial Analysis Intelligence Index (66, against 61 for GPT-6 Astra and 59 for Gemini 3.8 Flash) and the Coding Agent Index (70.4, against 67.0 and 61.1). But "best" and "right for you" differ: Astra is the only one that operates desktop software, and Gemini 3.8 Flash costs roughly a thirteenth as much.
Is GPT-6 Astra better than Claude Fable 5.1?
Not on general benchmarks — Fable 5.1 scores higher on both intelligence and coding. Astra is better when the task means operating a computer, doing frontier maths, or security work. They are the same list price, so choose on what the task actually requires rather than on cost.
How much cheaper is Gemini 3.8 Flash?
Roughly 13 times. Gemini 3.8 Flash lists at $0.75 per million input tokens and $3.75 output, against $10 and $50 for both GPT-6 Astra and Claude Fable 5.1. That pricing is promotional through 31 December 2026.
Which model is fastest?
Gemini 3.8 Flash, at roughly 305 tokens per second. GPT-6 Astra runs around 87 and Claude Fable 5.1 around 67 to 69. The speed gap between the cheapest and the most capable model is larger than the quality gap between them.
Should I pay for the expensive model?
Only if a human reads every output, or the cost of being wrong is high. Work out cost per finished task — price per run multiplied by average attempts, plus your review time — rather than price per million tokens. For high-volume work checked by a machine, the cheap model usually wins on that maths; for low-volume work checked by a person, it usually loses.
Can Claude Fable 5.1 or Gemini 3.8 Flash use my computer like Astra?
Not in the same way. Operating desktop applications directly — moving through menus, manipulating files across programs — is GPT-6 Astra's distinguishing capability in this comparison. OpenAI reports 72.6% on OSWorld 2.0, the benchmark for computer use, which also means roughly one task in four still fails.
Will this comparison still be accurate next month?
Partly. All three models launched in the first week of September 2026 and the leaderboard moved twice in that week alone. Treat the capability ordering as more durable than the exact scores, and re-check pricing before committing a budget — Gemini's promotional rate ends 31 December 2026.
Take twenty real tasks from your own work. Run them on Gemini 3.8 Flash and on Claude Fable 5.1. Count how many attempts each needed to produce something you would actually use.
That single afternoon will tell you more than every comparison table on the internet, because it measures your work rather than someone else's benchmark.
- Artificial Analysis, Intelligence Index and Coding Agent Index — the third-party measurements quoted throughout
- OpenAI, GPT-6 Astra — OSWorld 2.0 result and the Critical cyber classification
- The New Stack, OpenAI launches GPT-6 Astra
- Computerworld, Sam Altman calls the GPT-6 Astra rollout "messy"
- Jayant Solanki, Why the GPT-6 Astra benchmark records carry asterisks
Jayant Solanki
AI-Ready SEO, GEO & AIO strategist based in Indore, India. I compare tools on what they cost to finish a job, not on what they score, because the invoice and the leaderboard rarely agree.
More about Jayant →