GPT-5.6 vs Claude Opus 4.8 vs Fable 5: Which to Pick

Here is the decision before the argument. If you want the most capable single model and the budget is a second concern, pick Claude Fable 5. If you want the best capability you can reliably buy and run today at a sane price, pick Claude Opus 4.8. If you have API access to OpenAI’s new flagship and you are optimizing spend per finished task, GPT-5.6 Sol is the value play, because it nearly matches Fable 5’s intelligence at half the per-token rate and leads the field on agentic coding. For high-volume production work, Terra and Luna undercut everything Anthropic sells.

That is five models across two labs, and the interesting part is how tightly they cluster at the top. A year ago the frontier was a staircase. Now it’s a scrum.

The five models at a glance

Model	Price (in / out, per Mtok)	AA Intelligence Index	Context	Released
Claude Fable 5	$10 / $50	60	1M	Jun 9, 2026
GPT-5.6 Sol	$5 / $30	59	1.05M	Jul 9, 2026
Claude Opus 4.8	$5 / $25	56	1M	May 28, 2026
GPT-5.6 Terra	$2.50 / $15	55	1.05M	Jul 9, 2026
GPT-5.6 Luna	$1 / $6	51	1.05M	Jul 9, 2026

Prices come from each lab’s own rate card: OpenAI’s API pricing for the GPT-5.6 tiers, Anthropic’s Claude pricing for Opus 4.8, and the Fable 5 announcement for Fable. The intelligence scores are from the Artificial Analysis Intelligence Index, an independent aggregate benchmark, taken at each model’s maximum reasoning setting. Read the whole set beside the rest of the market on the AI model comparison page.

Two things jump out. First, four of these five models sit inside a five-point intelligence band. Second, the price spread inside that band is enormous: Luna costs about an eighth of Fable 5’s output rate for a score only nine points lower. The rest of this piece is about where in that band your work actually lives.

How these five were compared

The comparison holds one method across all five. Per-token prices are the published standard rates, not promotional windows or batch discounts. Capability uses the Artificial Analysis Intelligence Index and Coding Agent Index because they are run by a third party on the same tasks for every model, which is the only way a cross-lab number means anything. Where a lab’s own benchmark is the only source, it is labeled as vendor-reported, because a launch-day number from the company selling the model is a claim, not a measurement.

Cost per task is modeled, not scraped. An agentic coding job is a loop: read the repo, plan, edit, run tests, read the failures, retry. The bill is tokens per loop times loop count, with output weighted heaviest. That is the same method behind the AI coding cost calculator and the wider cost-per-task work on this site, so the numbers here line up with those.

Price against intelligence: the chart that settles most of it

Put price on one axis and measured intelligence on the other and the whole decision compresses into a picture.

Price vs intelligence: GPT-5.6 family, Opus 4.8, Fable 5

Item	Output price ($ per million tokens)	AA Intelligence Index
Claude Fable 5	$50	60
GPT-5.6 Sol	$30	59
Claude Opus 4.8	$25	56
GPT-5.6 Terra	$15	55
GPT-5.6 Luna	$6	51

Fable 5 is the most capable model in the set. It is also the only one priced at $10 / $50, twice Opus 4.8 and twice Sol on the input side. Anthropic positions it as its most capable model, shipped June 9, 2026, sitting a tier above Opus. The benchmark backs the positioning: 60 is the top of the board. The question is never whether Fable 5 is the best. It’s whether best-by-one-point is worth double the invoice.

Sol is the reason that question is live. At $5 / $30 it prices exactly half of Fable 5, and it scores 59 on the same index, one point back. Artificial Analysis, in its GPT-5.6 launch analysis, frames Sol as nearly matching Fable 5 at roughly one-third the cost per index task. Both framings are true and they measure different things: half is the per-token sticker gap, one-third is the cost of running the benchmark once the models’ token efficiency is folded in. Either way the direction is the same. Sol buys you the top of the frontier at a discount.

And Opus 4.8 is quietly the most sensible dot on the chart. It released May 28, 2026 at $5 / $25, the cheapest output rate of the three flagships, and scores 56. That is four points below Fable 5 for half the output price and no access drama. For most teams the honest recommendation is not the model that wins the benchmark. It’s this one.

Capability: who actually wins

Fable 5 wins on raw intelligence. By one point. On the AI model value leaderboard its composite scores are the highest in the set, which is exactly what you’d expect from a flagship priced like one.

But intelligence-in-aggregate is not the same as coding, and coding is where most of the money in this niche gets spent. Here the ranking flips.

AA Intelligence Index: five models compared

Item	Value
Claude Fable 5	60
GPT-5.6 Sol	59
Claude Opus 4.8	56
GPT-5.6 Terra	55
GPT-5.6 Luna	51

On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol scores 80 and leads the field. Fable 5 comes in at 77, tied with Terra, and Luna posts 75. So the cheaper OpenAI flagship is also the better agentic coder on the one cross-lab measure built for it. That is a genuine shift: for most of 2026 the assumption was that Anthropic owned coding outright, and on this index it no longer does. The launch-day detail sits in the GPT-5.6 Sol coding-leaderboard breakdown.

The vendor-reported numbers point the same way without settling it. On Terminal-Bench 2.1, OpenAI reports Sol at 91.91% in ultra mode and 88.76% in max mode, against Claude’s high-80s. Anthropic’s own SWE-bench Pro figures put Fable 5 in the low 80s. Both are lab-reported launch numbers, and at least one of Fable 5’s coding scores has been publicly contested by independent evaluators, so treat any single headline percentage as marketing until a third party reproduces it. The full set of tests, and which ones to trust, is on the AI benchmarks page.

What they are like to actually run

Benchmarks measure a model on someone else’s task. The more useful question is what these do on a real codebase over a real workday, and here the record is unusually well documented, because the launches came with named early users.

The strongest single account is from Cognition, the team behind the Devin coding agent. In an Anthropic case study published July 10, 2026, Cognition put Fable 5 at the top of FrontierCode, its internal benchmark that grades whether an agent’s output is actually mergeable, not just plausible. Fable 5 scored roughly 30% where the prior Opus generation scored around 10%, and it did so while running about eight hours of continuous autonomous work. Silas Alberti, Cognition’s SVP of Research, said the horizon was the shock: how long it stayed self-sufficient. He called it the first model he would trust to leave running overnight.

The biggest thing we noticed was the horizon: how long it can be self-sufficient. It was kind of a shocker, honestly.

Independent voices land in the same place. Simon Willison, who has tested every frontier release in public for years, pointed Fable 5 at his own open-source LLM library and watched it identify and implement four separate fixes, then ship a new release almost entirely written by the model. His word for it was “relentlessly proactive.” On Opus 4.8 a couple of weeks earlier, his reaction was different in tone and telling in its own way: it was, he wrote, refreshing to see an AI lab honestly describe a release as a minor incremental improvement. That is the gap between the two Anthropic models in a sentence: Fable 5 is the leap, Opus 4.8 is the dependable step.

The vendor testimonials rhyme, and should be read as testimonials, not measurements. Anthropic’s Fable 5 launch quoted Replit’s CTO Fabian Hedin saying it “one-shots” tasks that took a hundred prompts a year ago, and Mode Analytics reporting spreadsheet work running 25-30% faster. Anthropic also cited a migration inside Stripe’s roughly 50-million-line Ruby codebase completed in about a day; note the precise claim, which is one migration within that codebase, not a rewrite of all 50 million lines, a distinction Anthropic’s own wording supports and that early readers were quick to enforce.

GPT-5.6 has less of this on the record, because it went generally available only on July 9 and the independent production case studies have not landed yet. What it has instead is a loud practitioner reception centered on one theme: cost. Sam Altman framed Sol as a step change in dollars-per-task rather than a benchmark win, and that reframing, price as the product, is the honest way to read the whole GPT-5.6 launch.

What a hard task actually costs

Sticker rates undersell the gap, because output tokens dominate an agentic bill and the flagships burn a lot of them. Model one hard, multi-file task the same way for all five: roughly 60K input and 12K output tokens per loop, held at four loops each so no model gets credit for finishing faster. That isolates price from efficiency.

Modeled cost per hard task, equal loops

Tool	Cost per task	Multiple of baseline
Claude Fable 5	$4.80	-
GPT-5.6 Sol	$2.64	-
Claude Opus 4.8	$2.40	-
GPT-5.6 Terra	$1.32	-
GPT-5.6 Luna	$0.53	-

At equal loops, Fable 5 runs $4.80 against Opus 4.8’s $2.40: exactly double, because Fable 5’s rate is exactly double. Sol lands at $2.64, a hair above Opus despite the higher output rate, because its input side matches Opus at $5. Terra does the same job for $1.32 and Luna for $0.53.

Now let efficiency back in, because this is where Sol earns its reputation. Artificial Analysis measures cost per task on its Intelligence Index, not just per token, and there Sol runs about $1.04 against Fable 5’s $2.75, roughly a third. The reason is not the rate card; it’s that Sol finishes the same work while emitting far fewer output tokens, and output is the expensive half of the bill. OpenAI claims up to 54% fewer output tokens on agentic coding, and the AA numbers are consistent with a gap that large. This is the mechanism the sticker price hides: a model can charge a similar rate and still cost half as much, simply by being less verbose on the way to the answer.

Fable 5 has its own version of the same argument. Its case for the premium is that it finishes hard work in fewer loops, and Anthropic cites customer cases where it reaches an answer with far fewer tokens than the previous generation. If Fable 5 lands a task in three loops while a cheaper model needs five, the per-task gap narrows or closes. That is real, and it’s the whole argument for the flagship. But it’s your repository that decides it, not a press release. Drop your own token counts into the cost calculator and the crossover falls out. For anything routine, the loop counts converge and the cheaper model simply wins, which is the same conclusion the Fable 5 versus Opus 4.8 breakdown reached on Opus.

One cost trap deserves a flag. Sol ships an “ultra” mode that spins up parallel subagents, and it does lift scores: independent and vendor figures put Terminal-Bench 2.1 around 88.8% for standard Sol and 91.9% for ultra. But ultra spends roughly two to three times the tokens of a normal call, so a task that costs 15 cents in standard Sol can run 30 to 45 cents in ultra (figures relayed by secondary coverage, not a vendor rate card). The extra three points of benchmark are rarely worth triple the bill on everyday work. Reach for ultra on the genuinely hard problems and leave it off for the rest.

Access and the fine print

The spec sheet is not the whole story, because one of these models spent most of June switched off.

Fable 5 shipped on June 9. Three days later the US Commerce Department issued an export-control directive restricting access for foreign nationals, and because Anthropic could not verify nationality in real time, it disabled Fable 5 and Mythos 5 for everyone. What triggered it, per Anthropic’s own account of the redeployment, was a discovery by Amazon researchers: prompted to identify software vulnerabilities, Fable 5 could in at least one case produce code demonstrating how to exploit them. Anthropic says a new classifier now blocks that behavior in over 99% of cases, and the Commerce Department’s Center for AI Standards and Innovation tested the safeguards before access was restored.

It’s the first known use of US export-control authority against a specific commercial frontier model, and the full account is in the Fable 5 returns write-up. The practical read: Fable 5 is buyable again, but it carries a policy risk the others don’t.

There’s a second piece of Fable 5 fine print that will bite anyone wiring it into production. Fable 5 can refuse a request, and it does so in a way naive error handling misses. Per Anthropic’s refusal documentation, a refusal comes back as an HTTP 200 with stop_reason: "refusal" and an empty content array, so code that reaches for the first text block throws, and monitoring that only watches for 4xx and 5xx never sees it. Refusals fall into four categories (cyber, bio, frontier_llm, reasoning_extraction), and the cyber and bio filters can catch legitimate security or life-science work by proximity. Anthropic ships a server-side fallback that auto-retries a blocked Fable 5 call against Opus 4.8, and says more than 95% of Fable sessions never trip it, but “more than 95%” is not “never,” and the day it fires on you is the day you learn whether you handled it.

GPT-5.6 has its own access wrinkle, and it’s a geopolitical one. It previewed on June 26 to roughly 20 government-vetted partner organizations, API and Codex only, whose identities were shared with the government but never published, before going generally available on July 9 across ChatGPT, Codex and the API. OpenAI made its discomfort plain: it does not believe this kind of government access process should become the long-term default. For most of the window between the two labs’ launches, then, the frontier’s two strongest models were both partly or wholly ungettable, one for export policy and one for a preview gate. The launch specifics are in the Sol, Terra and Luna breakdown.

Opus 4.8 is the boring one, and in procurement boring is a feature. No suspension, no gate, no refusal-handling homework, stable rate card since the Opus 4.5 generation. Anthropic even pitches it on reliability: it says Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass unremarked, the kind of unglamorous improvement Simon Willison singled out as the honest sort. If you need to sign off on a model this quarter without a footnote about export policy or a new error path, that counts for something.

Which should you pick

For most engineering teams: Opus 4.8. It’s four points off the top of the intelligence board, it’s the cheapest flagship on output, and there’s no access asterisk. Pay for Fable 5 only when a specific hard problem justifies it.

For the hardest reasoning and long-horizon agentic work: Fable 5. It’s the most capable single model here, and if your tasks are genuinely at the edge of what models can do, one point of index score can be the difference between a solved ticket and a stuck one. Verify the loop-count savings on your own repo before you commit the whole team to double the rate.

For value on capability, if you have access: GPT-5.6 Sol. It matches Fable 5’s intelligence within a rounding error, leads it on agentic coding, and costs half as much per token. The only real caveat is whether you can get it and keep it, given the July rollout.

For high-volume production: Terra, then Luna. Terra delivers roughly GPT-5.5-class capability at $2.50 / $15, and Luna takes the floor at $1 / $6 for routine, well-scoped work where the capability gap doesn’t bite. Nothing Anthropic currently sells competes on price at these tiers.

Track prices and new releases as they land on the AI model release tracker, because in a market moving this fast the right answer has a short shelf life.

Frequently asked questions

Is GPT-5.6 better than Claude?

It depends on the model and the task. On the Artificial Analysis Intelligence Index, Claude Fable 5 leads at 60 with GPT-5.6 Sol one point behind at 59. On the Coding Agent Index, Sol leads at 80 versus Fable 5 at 77. So Claude Fable 5 edges general intelligence while GPT-5.6 Sol leads agentic coding, and Sol does it at half the per-token price.

Which is the cheapest of these models?

GPT-5.6 Luna at $1 per million input tokens and $6 per million output, followed by GPT-5.6 Terra at $2.50 / $15. Among the flagships, Claude Opus 4.8 is cheapest at $5 / $25, then GPT-5.6 Sol at $5 / $30, with Claude Fable 5 the most expensive at $10 / $50.

Is Claude Fable 5 worth double the price of Opus 4.8?

Only for the hardest work. Fable 5 scores 60 on the Artificial Analysis Intelligence Index against Opus 4.8 at 56, a four-point edge for exactly double the output rate. For routine and well-scoped tasks the loop counts converge and Opus 4.8 is the better spend. Reserve Fable 5 for problems where its capability cuts the number of attempts.

What are the three GPT-5.6 models?

Sol, Terra and Luna. Sol is the flagship at $5 / $30 per million tokens and leads the Artificial Analysis Coding Agent Index at 80. Terra is the balanced tier at $2.50 / $15. Luna is the budget tier at $1 / $6. All three went generally available on July 9, 2026 with a roughly 1.05M-token context window.

Which model should I pick for coding?

GPT-5.6 Sol leads the Artificial Analysis Coding Agent Index at 80, ahead of Claude Fable 5 at 77, so it is the strongest agentic coder on the cross-lab measure right now, at half Fable 5 per token. If you cannot access Sol, Claude Opus 4.8 is the strongest coder you can reliably buy and run today.

Why is GPT-5.6 Sol cheaper per task if the token rate is similar?

Because Sol emits far fewer output tokens to finish the same work, and output is the expensive half of the bill. OpenAI claims up to 54% fewer output tokens on agentic coding. On the Artificial Analysis Intelligence Index a task runs about $1.04 on Sol versus $2.75 on Fable 5, roughly a third, even though Sol per-token is half Fable 5 rather than a third.

Can Claude Fable 5 refuse requests?

Yes. Fable 5 ships safety classifiers that can refuse a prompt, returning an HTTP 200 with stop_reason "refusal" and an empty content array, so naive error handling can break. Refusals cover cyber, bio, frontier-model and reasoning-extraction categories. Anthropic offers a server-side fallback that auto-retries the call against Opus 4.8 and says over 95% of Fable sessions never trigger a refusal. Opus 4.8 does not carry these classifiers.

Sources

• OpenAI. (2026). API pricing. Primary; standard per-token rates for GPT-5.6 Sol ($5/$30), Terra ($2.50/$15) and Luna ($1/$6). Verified 2026-07-12. developers.openai.com/api/docs/pricing
• Anthropic. (2026). Introducing Claude Opus 4.8. Primary announcement; release date and positioning. Verified 2026-07-12. anthropic.com/news/claude-opus-4-8
• Anthropic. (2026). Claude Fable 5 and Claude Mythos 5. Primary announcement; Fable 5 pricing, release date, flagship positioning, export-control suspension. Verified 2026-07-12. anthropic.com/news/claude-fable-5-mythos-5
• Anthropic. (2026). Redeploying Fable 5. Primary; restoration of access after the directive was lifted. Verified 2026-07-12. anthropic.com/news/redeploying-fable-5
• Anthropic. (2026). Claude pricing. Primary; Opus 4.8 standard rates ($5/$25). Verified 2026-07-12. claude.com/pricing
• Artificial Analysis. (2026). Artificial Analysis Intelligence Index. Independent aggregate benchmark; intelligence scores at maximum reasoning (Fable 5 60, Sol 59, Opus 4.8 56, Terra 55, Luna 51) and the Coding Agent Index (Sol 80, Fable 5 77). Verified 2026-07-12. artificialanalysis.ai
• Artificial Analysis. (2026). GPT-5.6 has landed. Independent analysis; Sol nearly matching Fable 5 at roughly one-third the cost per index task ($1.04 vs $2.75), driven by lower output-token counts. Verified 2026-07-12. artificialanalysis.ai
• Anthropic / Claude. (2026). Working at the frontier: how Cognition trusts Claude Fable 5 to work through the night. Primary case study; FrontierCode ~10% to ~30%, ~8-hour autonomous runs, Silas Alberti quotes. Verified 2026-07-12. claude.com
• Willison, S. (2026). Claude Fable 5 and Claude Opus 4.8. Primary; independent practitioner testing (four fixes shipped as a release; the honest minor-release note). Verified 2026-07-12. simonwillison.net/2026/Jun/9/claude-fable-5, simonwillison.net/2026/May/28/claude-opus-4-8
• Anthropic. (2026). Handling refusals and fallbacks. Primary documentation; HTTP 200 refusal shape, four categories, Opus 4.8 fallback, >95% no-refusal figure. Verified 2026-07-12. platform.claude.com
• TechCrunch. (2026). OpenAI limits GPT-5.6 rollout after government request. Secondary; ~20 unnamed government-vetted preview partners, OpenAI position on government access. Verified 2026-07-12. techcrunch.com
• eesel AI. (2026). GPT-5.6 Sol ultra mode. Secondary; ultra-mode token multiplier and per-call cost range, relayed (not a vendor rate card). Verified 2026-07-12. eesel.ai
• TechJack Solutions. (2026). Claude Fable 5’s SWE-bench Pro score is contested. Secondary; independent evaluators dispute a vendor-reported coding figure. Verified 2026-07-12. techjacksolutions.com