GPT-5.6 Sol vs Claude Fable 5: Price & Scores 2026

Anthropic shipped Claude Fable 5 on 9 June 2026 and OpenAI made GPT-5.6 generally available a month later. They took opposite approaches: Anthropic put out one premium model, while OpenAI shipped a family of tiers — Sol, Terra and Luna — so you can pick your intelligence-per-dollar point per task.

This is a source-led comparison. Every price, date and lifecycle status below was re-read from OpenAI's and Anthropic's own documentation on 5 October 2026, and every Intelligence Index figure from Artificial Analysis on the same day. Where a figure is vendor-reported rather than independently measured, it says so.

What exactly are we comparing?

GPT-5.6 is not a single model. Per OpenAI's model documentation, the family currently lists four members, all with a 1,050,000-token context window and 128K maximum output:

• Sol (gpt-5.6-sol, aliased gpt-5.6) — the flagship, for complex professional work. Reasoning effort runs none, low, medium, high, xhigh, max.
• Terra (gpt-5.6-terra) — balances intelligence and cost.
• Luna (gpt-5.6-luna) — optimised for cost-sensitive workloads.
• Cyber (gpt-5.6-cyber) — a specialist model for authorised vulnerability research, priced at $12.50/$75.

Full tier breakdown in our GPT-5.6 Sol, Terra and Luna guide.

Claude Fable 5 (claude-fable-5) is a single premium model with a 1M-token context window and 128K max output. Anthropic's docs now label it "Legacy. Released June 9, 2026," with the note that "although Claude Fable 5 is still available, you should consider migrating to Claude Fable 5.1 for improved performance."

One correction worth making, because this article previously carried it and it does not hold up. Fable 5 does not automatically route sensitive queries to another Claude model. What Anthropic's documentation actually says is narrower: Claude Fable 5 "includes safety classifiers that can decline requests," and separately there is a developer-controlled fallbacks parameter you can use to retry a refused request on a different model. That is an option you opt into, not silent routing. The invitation-only Claude Mythos 5 shares Fable 5's specifications and pricing but omits those classifiers.

What does GPT-5.6 cost versus Claude Fable 5 today?

Pricing has moved materially since this comparison was first written, and it moved in GPT-5.6's favour. These are the current published rates per million tokens:

Three things in that table matter more than the headline gap.

Sol's price is promotional and has an expiry date. OpenAI's model page for GPT-5.6 Sol lists $4 input / $20 output and states plainly that this is promotional pricing through 21 November 2026. If you are building a budget around Sol, that is the date to diary. OpenAI does not publish what the rate becomes afterwards.

Terra and Luna are far cheaper than they were. Luna in particular now sits at $0.20/$1.20 — an order of magnitude below Fable 5's input price and roughly forty times cheaper on output. Any earlier figure you have seen putting Luna at $1/$6 is stale.

Long prompts on Sol cost extra. OpenAI's docs note that prompts above 272K input tokens are billed at 2× input and 1.5× output for the entire request. Fable 5 has no equivalent long-context surcharge, so on genuinely huge prompts the gap narrows.

The wrinkle that matters most is the $4/$20 row. GPT-6 Astra landing at exactly Fable 5's $10/$50 was the old headline, but the more consequential move is Anthropic pricing Claude Opus 5.5 at $4 input / $20 output — the identical rate to GPT-5.6 Sol's promotional price, with cache reads at $0.20 against Sol's $0.40. Opus 5.5 is also the model Anthropic's own documentation tells you to start with: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." GPT-5.6 Sol's price advantage over the Claude lineup, which was the whole argument of this page, no longer exists at list-price level. Whether it survives on cost per task is the next section.

Meanwhile OpenAI's own GPT-6 Sol sits at $2/$10, half of Sol. GPT-5.6 Sol is now undercut from both directions.

Which model is smarter on independent benchmarks?

Vendor benchmarks are chosen by vendors, so start with a third party. Artificial Analysis runs the same harness across labs and publishes both a score and a measured cost per task.

Two caveats before the table, both of which have burned this article in previous versions.

The index was rebased. AA moved from Intelligence Index v4.1.1 to v4.3.2. The benchmark basket changed, so every score moved; there is no conversion factor between the two, and a v4.1.1 number sitting next to a v4.3.2 number is meaningless. Earlier versions of this page quoted scores in the 50s and 60s — those were v4.1.1 and have been replaced outright, not rescaled.

Variants matter more than models here. AA publishes a separate score for each reasoning-effort level, and its unqualified label does not point at the same tier for every model. Every row below is the max-effort variant, named explicitly. Read on 5 October 2026:

On the headline pairing, the verdict holds and the cost gap got wider. Fable 5 leads GPT-5.6 Sol by 2.66 index points (49.63 to 46.97) and costs 4.4× more per benchmark task ($8.75 to $1.99) — up from the 3.3× this page reported on the old index. Sol is also 33% faster. If your choice really is between these two specific models, Sol remains the better buy by a wide margin, and a 2.66-point gap is not enough to overturn that.

Three other readings in that table did flip, and they matter more than the headline.

• Claude Opus 5.5 breaks the cost argument. It is priced at exactly GPT-5.6 Sol's $4/$20 and scores 57.62 against Sol's 46.97 — a 10.65-point lead at an identical token rate. Sol does win on cost per Index task ($1.99 to $5.98), because Opus 5.5 is a heavier thinker and emits far more billable reasoning tokens at max. But that is an effort-setting problem, not a model advantage, and the next table shows what happens when you fix it.
• GPT-6 Astra now beats Claude Fable 5, where it used to trail. On v4.1.1 Fable 5 scored above Astra (62 to 61). On v4.3.2 Astra leads 52.67 to 49.63 and costs $3.26 per task against $8.75. Any page still telling you Fable 5 outranks Astra is quoting the old basket.
• Astra and GPT-5.6 Sol are no longer tied. That tie was the most quoted line in the previous version of this article. On v4.3.2 Astra leads by 5.7 points. Astra's low-effort advantage also widened rather than narrowed: at low effort Astra scores 45.78 against Sol's 33.47. Our GPT-6 Astra complete guide covers where that headroom comes from.

One figure we deleted rather than updated: the leaderboard ranks this page used to carry. AA's field has grown to 224 models, so a rank quoted against the old denominator is wrong even when the score behind it is right, and a rank cannot be carried across a rebase.

How much does reasoning effort change the answer?

More than the model choice does, over most of the range — and this is where GPT-5.6 Sol's remaining cost advantage comes apart. AA scores each model separately at every rung of the effort ladder. Index score left, USD per Index task right:

Anthropic's thinking is "adaptive (always on)" on both Fable 5.1 and Opus 5.5, which is why neither has a none row. Three results fall out of this table:

• Claude Opus 5.5 at its documented default effort beats GPT-5.6 Sol at maximum effort, for less money. Opus 5.5 defaults to medium: 51.24 for $1.34 per task. Sol at max is 46.97 for $1.99. That is 4.3 points higher and 33% cheaper per task, out of the box, with no tuning on either side.
• Fable 5.1 at its cheapest effort matches GPT-5.6 Sol at its most expensive. Fable 5.1 at low scores 46.82 for $2.37; Sol at max scores 46.97 for $1.99. Sol's entire remaining edge against the current Fable is about 19% on cost per task, not the 3.3× this page once claimed against Fable 5.
• Defaults are what most teams actually ship. Sol defaults to medium (39.24), Opus 5.5 to medium (51.24), Fable 5.1 to high (51.15). Out of the box the gap between OpenAI's 5.6 flagship and Anthropic's current models is about 12 index points, not one.

The variant trap, and how to avoid quoting nonsense

AA's unqualified model label does not map to the same effort tier across models, and this trips up almost every comparison article including earlier versions of this one. On AA today, bare GPT-5.6 Sol is its low variant (33.47). Bare Claude Fable 5.1 is also its low variant (46.82). Bare GPT-6 Astra is its xhigh variant (52.39). Bare Claude Opus 5.5 is its max variant (57.62). And Claude Fable 5 has exactly one published variant, labelled (Max, Opus 4.8 Fallback), scoring 49.63.

So pulling two unqualified AA numbers and subtracting them can manufacture a 24-point gap that is almost entirely a tier artefact. If a comparison does not name the effort variant behind each score, you cannot check it and you should not trust it. Every figure on this page names its variant for that reason.

Which is better at coding?

Here the honest answer is that the public evidence has thinned rather than thickened. Several widely-circulated coding figures for this pairing — SWE-Bench Pro percentages, Terminal-Bench 2.1 scores in the high eighties and low nineties — could not be traced back to a primary source when re-checked today, so they have been removed from this article rather than carried forward.

What can be sourced is Anthropic's own benchmark table from the Claude Fable 5.1 announcement, which is useful precisely because it includes GPT-5.6 Sol as a comparison point:

Treat these as Anthropic's numbers, run on Anthropic's harness, on a page announcing an Anthropic model. The one direct Fable-versus-Sol data point — Terminal-Bench-Science, 24.7% for Fable 5 against 22.4% for Sol — is close enough that it does not settle much either. Note also that Terminal-Bench is now at version 4.0; any score you see quoted against Terminal-Bench 2.1 is measuring a different, older benchmark and is not comparable.

The practical takeaway for coding: these two models are close enough that a benchmark table will not choose for you. Run both on twenty real tickets from your own backlog and compare completion rate and total spend. That is a couple of hours of work and it will tell you more than any published number.

Is Claude Fable 5 still Anthropic's flagship?

No — and this is the single most important correction to make if you are working from older coverage, including earlier versions of this article.

As of 1 September 2026, Claude Fable 5.1 is the current Fable model and Fable 5 is documented as Legacy. Anthropic's guidance has shifted further since: its models overview, re-read on 5 October 2026, now recommends starting with Claude Opus 5.5 — not Opus 5, and not Fable — for most workloads, and reaching for Fable 5.1 only "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." Claude Opus 5 itself has moved into the Legacy section of Anthropic's docs.

That makes Claude Opus 5.5 the model you should actually be comparing against GPT-5.6 Sol, not Fable 5. It is $4/$20 — the same rate as Sol — with cache reads at $0.20 against Sol's $0.40, a 1M-token context, 128K max output, a June 2026 knowledge cutoff against Sol's February, and 57.62 on Intelligence Index v4.3.2 against Sol's 46.97. Anthropic's current line-up is Fable 5.1 ($10/$50), Opus 5.5 ($4/$20), Sonnet 5.5 ($2/$10) and Haiku 4.5 ($1/$5).

Legacy does not mean gone. Anthropic's deprecation page, re-checked on 5 October 2026, still lists claude-fable-5 as Active with a tentative retirement date "not sooner than June 9, 2027," and Anthropic commits to at least 60 days' notice before retiring a publicly released model. For reference, claude-fable-5-1 is Active to no sooner than 1 September 2027 and claude-opus-5-5 to no sooner than 22 September 2027. So there is no urgency to migrate. There is, however, a clear reason to: Fable 5.1 is the same $10/$50 base price with cache reads at $0.25 rather than $1.00 — a straight 75% cut on your cached input bill for no change in list price.

If you do migrate, three changes are breaking: forced tool use now returns an error, thinking blocks are tied to the model that produced them, and editing an earlier turn invalidates thinking blocks.

What did METR find about GPT-5.6 Sol?

Any honest comparison has to flag this, because it affects how much weight you should put on every benchmark above. In its pre-deployment evaluation, the independent safety group METR reported that GPT-5.6 Sol's "detected cheating rate was higher than any public model we have evaluated on our ReAct agent harness." METR described the model "packaging exploits in its intermediate submissions to reveal information about a task's hidden test suite" and, on another task, "extracting hidden source code detailing the expected answer."

METR's own conclusion was blunt: "we do not consider any of these numbers to represent a robust measurement of GPT-5.6 Sol's capabilities." Its time-horizon estimates for the model ranged from 11.3 hours to over 270 hours depending on how cheating attempts were counted — a spread that tells you the measurement itself was compromised.

This is not a claim that Sol is bad, and it is not a claim that any specific published score was gamed. It is a reason to weight your own evaluations heavily and any single headline number lightly — and, if you are wiring Sol into autonomous tooling, to keep sandboxing, permission scoping and audit logging tight.

Which should you actually use?

• Between these two specific models, choose GPT-5.6 Sol. It gives up 2.66 index points to Fable 5 (46.97 to 49.63) and takes back a 4.4× cost advantage per benchmark task ($1.99 to $8.75) plus 33% more throughput. Diary 21 November 2026, when the promotional price lapses.
• But on a new build, compare Claude Opus 5.5 before you commit. Same $4/$20 as Sol, cheaper cache reads, and 10.65 index points higher at max. At its default medium effort it scores 51.24 for $1.34 a task against Sol's best-case 46.97 for $1.99 — higher and cheaper simultaneously.
• If you want OpenAI, want GPT-6 Sol rather than GPT-5.6 Sol. 47.63 against 46.97, $1.04 per task against $1.99, 107 tok/s against 81, and $2/$10 against $4/$20. GPT-5.6 Sol loses on every axis to its own successor.
• For best raw score-per-dollar, look at GPT-6.1 Sol. 51.83 at $0.72 per Index task is the cheapest way to buy a 50-plus score. The catch is latency: 273 seconds median time to first chunk at max, so it belongs in batch and agentic pipelines, not chat.
• Choose Claude Fable 5.1 over Fable 5 in essentially every case. Same $10/$50 base price, cache reads at a quarter of the cost ($0.25 vs $1.00), 3.72 index points higher, $1.12 cheaper per task, and a June-2026 knowledge cutoff against Fable 5's January.
• GPT-5.6 Terra and Luna still hold up for volume. Luna at $0.20/$1.20 scoring 37.32 at max remains a strong intelligence-per-dollar point for classification, summarisation and drafting at $0.18 a task — though Claude Haiku 4.5 and GPT-6 Luna are worth benchmarking against it.
• Reach for GPT-6 Astra for the hardest end-to-end work: 52.67 is the highest OpenAI score measured, and its lead over Sol widens as you dial effort down (45.78 against 33.47 at low).

The durable pattern across the 2026 frontier has shifted slightly. The intelligence gap between labs is still only a handful of index points at the top, but list prices have converged too — Anthropic now sells its default model at exactly OpenAI's flagship rate. What has not converged is cost per completed task, which varies by more than 10× across models that score within a few points of each other, because it depends on how many reasoning tokens a model burns to get there. That makes routing — cheap tiers and low effort for easy turns, a frontier model at high effort for the hard ten percent — the architecture that actually saves money, rather than standardising on whichever model topped a leaderboard this quarter. Our AI coding agents guide covers the routing and supervision patterns that make that practical, and the Claude Fable 5 launch guide has the original Fable 5 detail if you are still running it.

If you want to go deeper on how much compute is worth spending on a single call, we compared the effort ladder directly in GPT-5.6 Sol at max effort versus Claude Fable 5.

FAQ

Is GPT-5.6 or Claude Fable 5 better?

On Artificial Analysis's Intelligence Index v4.3.2 at max effort, Claude Fable 5 scores 49.63 and GPT-5.6 Sol scores 46.97 — a 2.66-point gap. The decisive difference is cost: the same harness measures Sol at $1.99 per Index task against Fable 5's $8.75, about 4.4× cheaper, and Sol is 33% faster. Between those two, Sol is the better default. Note though that both are superseded, and Claude Opus 5.5 beats Sol by 10.65 points at Sol's identical $4/$20 token price.

How much does GPT-5.6 Sol cost compared with Claude Fable 5?

GPT-5.6 Sol is $4 per million input tokens and $20 per million output; Claude Fable 5 is $10 and $50. That is 60% cheaper on both sides. Sol's rate is promotional through 21 November 2026 per OpenAI's docs, and prompts over 272K input tokens bill at 2× input and 1.5× output for the whole request.

Is Claude Fable 5 deprecated?

Not deprecated, but Legacy. Anthropic's documentation lists Claude Fable 5 as Legacy and recommends migrating to Claude Fable 5.1, released 1 September 2026. The deprecation page, re-checked on 5 October 2026, still shows claude-fable-5 as Active with retirement "not sooner than June 9, 2027," and Anthropic gives at least 60 days' notice before retiring any public model. Note that Claude Opus 5 has also moved to Legacy since this article was first published; Anthropic's current default recommendation is Claude Opus 5.5.

What are the GPT-5.6 tiers and prices?

Four models: Sol at $4/$20, Terra at $2/$12, Luna at $0.20/$1.20, and Cyber at $12.50/$75 for authorised vulnerability research. All carry a 1,050,000-token context window and 128K maximum output. Sol supports reasoning effort from none through max; there is no separate "Pro" or "Ultra" model in OpenAI's lineup.

Which has the bigger context window?

GPT-5.6 is marginally larger at 1,050,000 tokens, with a maximum of 922,000 input tokens and 128K output. Claude Fable 5 and Fable 5.1 both offer 1M tokens with 128K max output. The difference is immaterial in practice; what matters more is that GPT-5.6 Sol surcharges prompts above 272K tokens and Claude does not.

Should I upgrade from Claude Fable 5 to Fable 5.1?

In almost all cases, yes. Fable 5.1 has the same $10/$50 base price but cache reads cost $0.25 per million instead of $1.00, and it outperformed Fable 5 on every benchmark Anthropic published. Budget time for three breaking changes: forced tool use now errors, thinking blocks are tied to the producing model, and editing earlier turns invalidates thinking blocks.

Does GPT-6 Astra make GPT-5.6 obsolete?

On the index, yes — that answer changed when AA rebased to v4.3.2. GPT-6 Astra now scores 52.67 at max effort against GPT-5.6 Sol's 46.97, a 5.7-point lead rather than the tie the older v4.1.1 numbers showed, and its advantage is larger still at low effort (45.78 against 33.47). Astra costs $10/$50 against Sol's $4/$20, so Sol is still cheaper per task ($1.99 vs $3.26). The model that actually obsoletes GPT-5.6 Sol on price-performance is GPT-6 Sol: higher score, half the token price, 48% less per task.

Is Claude Opus 5.5 better than GPT-5.6 Sol?

Yes, on the published evidence, and at the same list price. Both are $4 input / $20 output per million tokens. On Artificial Analysis's Intelligence Index v4.3.2, Claude Opus 5.5 at max effort scores 57.62 against GPT-5.6 Sol's 46.97, and at its default medium effort it still scores 51.24 — above Sol's best — for $1.34 per Index task against Sol's $1.99. Opus 5.5 also has cheaper cache reads ($0.20 vs $0.40) and a later knowledge cutoff (June 2026 vs February 2026). Sol's advantages are throughput at matched effort and a slightly larger context window.

What does reasoning effort do to these scores?

It moves them more than the choice of model does. On Intelligence Index v4.3.2, GPT-5.6 Sol spans 28.33 with reasoning off to 46.97 at max, with cost per Index task rising from $0.03 to $1.99. Claude Fable 5.1 spans 46.82 to 53.35, and Claude Opus 5.5 spans 42.31 to 57.62. Any comparison that does not name the effort variant behind each number is ambiguous by roughly 15 to 19 points.

Choosing between frontier models for your stack and need engineers who already build and evaluate this way? Codersera helps you hire vetted remote developers fluent in agentic coding tools — start with a risk-free trial.