Codex (GPT-5.5) vs Claude Fable 5 for coding: which is actually worth paying for?

Update — June 30, 2026: Claude Fable 5 (claude-fable-5) is back. After the June 12 US export-control directive that had taken both models offline was lifted on June 30, Anthropic is redeploying Fable 5 globally on July 1, 2026 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork (cloud providers to follow). It returns with new safety classifiers that block certain harmful cybersecurity tasks; on Anthropic’s own surfaces a blocked request falls back to Opus 4.8, and Anthropic flagged more false positives in routine coding and debugging that it will keep refining. See what changed in the redeployment and how to handle the classifier fallback as a developer. The analysis below describes Fable 5 as it launched; treat its availability and pricing as current again.

What are we actually comparing here?

Let me clear up the naming first, because it trips people up. When folks say “Codex 5.5,” they mean OpenAI’s GPT-5.5 model running inside Codex — OpenAI’s agentic coding tool and CLI. There is no separate “Codex 5.5” model. It’s GPT-5.5 doing the work; Codex is the harness around it.

GPT-5.5 shipped on April 23, 2026, and you can reach it in ChatGPT, in Codex (Plus/Pro/Business/Enterprise), and through the API. On the other side, Claude Fable 5 landed on June 9, 2026. It’s a new “Mythos-class” tier that sits above Opus — the safeguarded version of Mythos — and it launched at #1 on the Artificial Analysis Intelligence Index.

Here’s the lens I bring to this as someone who ships AI features in production: the interesting question is not “which one scores higher on a leaderboard.” It’s “which one gives me the lowest cost per actually-solved task.” A model that wins every benchmark but costs you double on work the cheaper model already handles is not a win — it’s a leak in your budget. So I’m going to weigh capability against price the whole way through.

Who wins the benchmarks — and how big is the gap?

I’ll be straight: on raw capability, Fable 5 is clearly ahead. The headline number comes from Anthropic’s own Fable 5 announcement — SWE-Bench Pro: Fable 5 at 80.3% vs GPT-5.5 at 58.6%. That’s a roughly 22-point gap (+21.7), and SWE-Bench Pro is about resolving real, end-to-end GitHub issues, which is the closest thing we have to “did it actually fix the codebase.” Fable is comfortably in front there.

The rest of Anthropic’s own announced gaps point the same direction: legal reasoning 13.3% vs 2.1%, Fable spatial reasoning at 38.6%, vision at state-of-the-art, and FrontierCode roughly 2x Opus 4.8. Third-party comparisons report more of the same shape — Terminal-Bench around 88% for Fable vs ~83% for GPT-5.5 on a Codex CLI run (OpenAI separately reported GPT-5.5 at ~82.7% on Terminal-Bench 2.0), and GraphWalks long-context at 1M reportedly 68.1% vs 45.4%. Treat those exact third-party figures as soft; the safe, headline comparison is SWE-Bench Pro plus the pricing.

The headline capability gap: SWE-Bench Pro

Fable 580.3%

GPT-5.558.6%

So on capability, Fable wins, and it isn’t close on the headline number. But a benchmark lead is not the same as best value — and that’s the whole reason this post exists.

What does each one actually cost to run?

Here’s where the math turns. Verified list pricing, per million tokens:

• GPT-5.5: $5 input / $30 output (Standard). GPT-5.5 Pro: $30 / $180.
• Claude Fable 5: $10 input / $50 output — roughly 2x GPT-5.5, and roughly 2x Opus 4.8.
• Claude Opus 4.8 (reference): $5 / $25.

Fable does offer ~$1/M cached input with its 90% prompt-caching discount, which helps on repeated context. But the sticker is the sticker: Fable is about twice the price per token.

Price per 1M tokens (input / output)

GPT-5.5 input$5

GPT-5.5 output$30

Fable 5 input$10

Fable 5 output$50

Opus 4.8 input$5

Opus 4.8 output$25

Now the part that flips the value math. GPT-5.5 uses ~40% fewer output tokens than GPT-5.4 on the same Codex tasks, and it matches GPT-5.4 per-token latency. It’s notably token-efficient — it gets to the answer with less verbosity. Combine that with being half the per-token price, and the gap in real spend gets wider than the price list alone suggests.

On context, both reach 1M tokens — but with a catch. GPT-5.5 is 1M via API and 400K inside Codex; Fable 5 is the full 1M for text, image, and file input. If you’re feeding enormous context through the tool itself, that 400K cap is a real constraint worth knowing before you commit.

The combined effect: half the per-token price and fewer output tokens means GPT-5.5 is often the lower cost-per-successful-task option at volume. I dug into this kind of math more in how much it costs to add AI to your app.

Cost-per-solved-task: where the 2x premium actually pays off

So we’ve got the tension stated plainly: Fable’s 22-point SWE-Bench Pro lead versus GPT-5.5 being half the price and ~40% more output-efficient. How do you reconcile that?

My read: for high-volume and terminal-style coding, GPT-5.5 usually wins cost-per-successful-task. It’s cheaper per token and it spends fewer tokens, so on the broad middle of everyday coding — where both models solve the task anyway — you’re paying double for an outcome you already had.

Fable 5 earns its 2x on the hard tail: the jobs it actually finishes that GPT-5.5 can’t. That’s long-horizon autonomy and codebase-wide resolution — exactly where that SWE-Bench Pro gap shows up. The concrete proof of this shape is in Anthropic’s own announcement: Stripe ran a 50-million-line Ruby migration in one day with Fable 5. That is not a task you’d casually hand to a cheaper model and hope.

The cost-per-task tradeoff (and one billing gotcha)

GPT-5.5 price$5 / $30 per 1M

Fable 5 price$10 / $50 per 1M (~2x)

Output efficiencyGPT-5.5 ~40% fewer output tokens

SWE-Bench ProFable 80.3% vs GPT-5.5 58.6%

Fallback billingFlagged topics route to Opus 4.8 — billed $5 / $25

So the reframe I’d offer: don’t pay 2x for a model on tasks the cheaper model already solves. Pay 2x only where capability converts — where it turns into a finished task that otherwise wouldn’t have shipped. If you’re still deciding whether you even need the most powerful model for your use case, I’d start with AI agents vs automation tools before reaching for the priciest option.

The Fable fallback and retention caveats you need to know

Two verified caveats on the Fable side that affect real billing and compliance — fair warning, not a hit piece.

First, the safety fallback. Fable 5 auto-routes cybersecurity, biology, and distillation queries to Claude Opus 4.8 and notifies you when it does. The billing nuance: on those flagged topics you effectively get Opus 4.8, and you’re billed at Opus 4.8 rates ($5 / $25), not Fable rates. The scope is narrow — Anthropic says over 95% of sessions never fall back — so this is an edge case, not a dealbreaker. But if your work genuinely touches those domains, you should know your “Fable” requests may not be Fable.

Second, and this one matters more for regulated teams: a mandatory 30-day data-retention policy applies to all Fable/Mythos traffic — even customers who hold zero-retention agreements. My honest opinion is that the retention policy weighs heavier than the fallback for anyone in a compliance-sensitive shop. Factor it into your default-model decision, not just an afterthought. I get into the price-plus-retention question more in is Claude Fable 5 worth it.

So which do I actually bill each task to?

Route by task shape, not by leaderboard ranking. That’s the whole discipline.

Which model do I bill this task to?

1. 01Everyday terminal coding?GPT-5.5 in Codex (or Opus 4.8) — cost-efficient and token-efficient default.
2. 02Hardest long-horizon or codebase-wide job?Claude Fable 5 — where the 2x premium converts into a solved task (the Stripe migration shape).
3. 03Accuracy-critical deep dive?GPT-5.5 Pro — when being right is worth the price.

• Default: GPT-5.5 (or Opus 4.8) for everyday terminal-centric coding. Cheap, fast, efficient.
• Hard tail: Claude Fable 5 for the hardest long-horizon, codebase-wide work where the premium actually converts.
• Accuracy-critical: GPT-5.5 Pro for deep dives where being right beats being cheap.

My decisive pick? GPT-5.5 in Codex is my daily driver, and I keep Fable 5 on a leash for the jobs that genuinely need it. I route — I don’t crown one winner.

Frequently asked questions

Is “Codex 5.5” a separate model? No. It’s GPT-5.5 running inside OpenAI’s Codex agent. The model is GPT-5.5; Codex is the tool.

Is Fable 5 worth 2x the price? Only for the hardest long-horizon and codebase-wide jobs. For everyday coding, GPT-5.5 is usually the better cost-per-task — it’s half the price and ~40% more output-efficient. (That’s my opinion, grounded in the verified pricing and the SWE-Bench Pro gap.)

Which has the bigger context window for coding? Both offer 1M tokens, but Codex caps GPT-5.5 at 400K in-tool, while Fable 5 gives you the full 1M. For massive single-context jobs, Fable has the edge.

Does Fable 5 ever silently downgrade? Not silently — it auto-routes cybersecurity, biology, and distillation queries to Opus 4.8 and notifies you. Over 95% of sessions never fall back.

What about data retention? Fable/Mythos traffic carries a mandatory 30-day retention, even under zero-retention agreements. Worth knowing before you make it your default.

The bottom line

Fable 5 is the more capable model — the SWE-Bench Pro number (80.3% vs 58.6%) is real and it’s large. GPT-5.5 in Codex is usually the better value: half the price, fewer output tokens, a strong cost-efficient default. Pay for capability only where it converts into tasks that otherwise wouldn’t ship.

Both are genuinely good. The win here is routing, not loyalty — and if you want more of how I think about this, the rest of my AI engineering guides go deeper. Don’t take my word on the numbers: check them yourself at OpenAI’s GPT-5.5 announcement and Anthropic’s Claude Fable 5 announcement.