Claude Fable 5 vs GPT-5.5: Benchmarks & Cost Compared

By mid-2026, the choice for most teams building serious AI features comes down to two models: Anthropic's Claude Fable 5 — released June 9 as the generally available half of a two-model launch — and OpenAI's GPT-5.5, which shipped April 23 and is now the model a large share of engineering teams already run inside Codex. This is the head-to-head: who wins on capability, who wins on cost, and which one fits the work you are actually trying to ship.

The short version. On Anthropic's own published benchmark table — the only one that pits the two directly — Fable 5 leads GPT-5.5 on every row, often by wide margins. But GPT-5.5 costs half as much per token, ties or beats Opus-class models on terminal coding through its own Codex harness, leads on parts of long-context retrieval and abstract reasoning, and carries an enormous deployed ecosystem. And in a telling convergence, both labs now treat cybersecurity and biology as high-risk and gate those capabilities behind vetted-access programs.

For the full picture on each model, see our Claude Fable 5 & Mythos 5 release breakdown and our GPT-5.5 complete guide. This post is the comparison, with the benchmark caveats stated plainly and a routing guide at the end.

These models come from different release strategies. GPT-5.5 shipped on April 23, 2026 (API access the next day) as OpenAI's single flagship, positioned around Codex and computer use — a model built to take a messy, multi-part task and carry it across tools until it is finished. By June it is mature, widely deployed, and the default for a large share of engineering teams.

Claude Fable 5 shipped seven weeks later, on June 9, 2026, as the generally available half of a two-model launch. Its twin, Mythos 5, is the same underlying model with safeguards lifted, restricted to vetted partners. Fable 5 is the newer model on paper, and Anthropic's benchmarks reflect that — but newer is not automatically better for your use case, which is the entire point of running the comparison rather than reading the headline.

The honest framing: Fable 5 is the capability leader on the published numbers, and GPT-5.5 is the value-and-ecosystem play that has had two months to embed itself in real workflows. For the deeper Opus-versus-GPT history that informs this matchup, see our Claude Opus 4.8 vs GPT-5.5 comparison.

The table below is Anthropic's published comparison — the only source that benchmarks Fable 5 and GPT-5.5 under the same conditions. Read it as Anthropic-reported. The reassuring part: where these numbers overlap with OpenAI's own published table (SWE-Bench Pro 58.6%, OSWorld-Verified 78.7%, Humanity's Last Exam 41.4% / 52.2%, Terminal-Bench via Codex CLI 83.4%), the two labs agree. Starred (*) rows are where Anthropic shows the restricted Mythos 5 score — the Fable 5 you can deploy falls back toward Opus 4.8 there.

Source: Anthropic Claude Fable 5 / Mythos 5 benchmark table, June 9, 2026. Starred (*) rows show the restricted Mythos 5 score; on those, the deployable Fable 5 performs closer to Claude Opus 4.8 because of its safeguard fallbacks. GPT-5.5's Terminal-Bench 2.1 figure is from its own Codex CLI harness and is not directly comparable to a public-harness run.

A capability table that Fable 5 sweeps can make GPT-5.5 look beaten. It is not. GPT-5.5 wins the arguments that decide most real deployments.

Price, decisively. At $5/$30 per million tokens, GPT-5.5 is half the cost of Fable 5's $10/$50. For high-volume production traffic, that gap compounds into real money, and it means the capability difference has to actually matter for your task before the premium is worth paying.

Terminal coding and the Codex loop. Terminal-Bench 2.0 at 82.7% is GPT-5.5's headline, and its Codex CLI delivers 83.4% on Terminal-Bench 2.1. OpenAI reports more than 85% of its own staff use Codex weekly, and the model is tuned to finish the same Codex tasks with fewer tokens than GPT-5.4. If your team already lives in Codex, that integration is a real, compounding advantage.

Abstract reasoning and research math. OpenAI reports GPT-5.5 at 85.0% on ARC-AGI-2 and 35.4% on FrontierMath Tier 4 (39.6% for GPT-5.5 Pro) — strong results on the hardest reasoning and research-math evals, areas Anthropic's Fable 5 table does not directly cover.

Long-context retrieval, on its own benchmarks. GPT-5.5 posts 74.0% on OpenAI's MRCR v2 at 512K-1M tokens, a genuine long-context strength. The honest caveat is in the next section: on GraphWalks, Claude leads. The two labs win different long-context evals, so this is a strength for GPT-5.5, not a settled victory.

GPQA near-parity. On GPQA Diamond, GPT-5.5 scores 93.6% — within noise of the near-saturated frontier (Opus 4.7 94.2%, Gemini 3.1 Pro 94.3%). Anthropic did not include GPQA in the Fable 5 table, so there is no direct figure to beat, but GPT-5.5 is clearly competitive on graduate-level STEM reasoning. For the deeper cost modeling behind all of this, see our agentic coding cost analysis.

Fable 5's advantages cluster around the hardest, longest, most multi-step work — exactly where the extra capability justifies the extra cost.

Codebase resolution. The 80.3% vs 58.6% SWE-Bench Pro gap is the headline. For end-to-end resolution of real GitHub issues — read the codebase, find the fix, carry it through related files, not just operate a terminal — Fable 5 is in a different tier. On the harder FrontierCode Diamond set the gap is starker still: 29.3% vs 5.7%.

Knowledge work and computer use. GDPval-AA 1932 vs 1769, OSWorld-Verified 85.0% vs 78.7%, AutomationBench 17.4% vs 12.9%, and a Legal Agent Benchmark gap of 13.3% vs 2.1%. For multi-stage analysis, document-heavy work, and agents that operate real software, Fable 5 leads consistently.

Long-context resolution, on GraphWalks. Where GPT-5.5 leads MRCR, Claude leads GraphWalks: at 1M tokens, Opus 4.8 scores 68.1% on GraphWalks BFS against GPT-5.5's 45.4% (Anthropic-reported). At those retrieval rates, the weaker model can miss roughly one in five facts the stronger one finds at the same context size. If your retrieval pattern looks like graph traversal over a large corpus, that gap matters.

Long-horizon autonomy. Fable 5 is built to run for days, plan across stages, delegate to sub-agents, and validate its own work at high effort. That maps directly onto the kind of sustained agentic delivery — large migrations, multi-day build-and-test loops — where a higher per-token price is dwarfed by the senior hours it replaces. See the full capability picture in our Fable 5 release breakdown.

Capability decides whether a model can do the job; price decides whether you can afford to run it at scale. Here is the rate card for both, with Opus 4.8 included because it is Fable 5's own safeguard fallback and undercuts both on output.

The per-task math. Take a representative 100K-input / 20K-output task. GPT-5.5: $0.50 input + $0.60 output = about $1.10. Fable 5: $1.00 input + $1.00 output = about $2.00 — roughly 80% more before caching. Prompt caching narrows the input side (Fable 5's 90% input discount on repeated context is strong), but the $50 vs $30 output rate keeps GPT-5.5 ahead on any output-heavy workload.

Where the math flips. Above 272K input tokens in a session, GPT-5.5 applies a surcharge — 2x input and 1.5x output — to the entire session, not just the tokens over the line. Fable 5 has no published long-context surcharge. So for genuinely large-context work (whole-repository passes, long document sets), GPT-5.5's headline price advantage erodes and Fable 5 becomes more competitive than the rate card first suggests. Model your real token distribution — the crossover point is workload-specific. Our agentic coding cost breakdown walks through the methodology.

There is no single winner, and standardizing on one model for everything leaves value on the table. The disciplined approach is to route by task shape — and, for anything high-stakes, to test both on your own pipeline before committing. Here is the practical map.