Codex (GPT-5.5) vs Claude Fable 5: A Builder's Honest Head-to-Head
Claude Fable 5 leads on the hardest codebase-wide and long-horizon coding (SWE-Bench Pro 80.3% vs 58.6%), but costs about twice as much ($10/$50 vs $5/$30). GPT-5.5 in Codex is the cost-efficient, token-lean default. Route by task: Fable for the hard tail, GPT-5.5 otherwise.
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.
Wait, what exactly is “Codex 5.5”?
Let me clear up the naming, because it trips people up. “Codex 5.5” isn’t a model. It’s OpenAI’s GPT-5.5 running inside Codex — Codex being OpenAI’s agentic coding tool/CLI. GPT-5.5 shipped on April 23, 2026, and it’s available in ChatGPT, in Codex (Plus/Pro/Business/Enterprise), and in the API.
On the other side: Claude Fable 5, which Anthropic released on June 9, 2026. Fable 5 is a new “Mythos-class” tier that sits above Opus — it’s the safeguarded version of Mythos, and it launched at #1 on the Artificial Analysis Intelligence Index.
I build AI features and agents in production, so I care about one thing here: which one do I point at which task, and what does it cost me to be right. This is a fair head-to-head, not a fanboy post and not a hit piece. Both models are genuinely good. They’re good at different things.
The benchmarks: who actually codes better?
The headline number — verified, straight from Anthropic’s own Fable 5 announcement — is SWE-Bench Pro: Fable 5 at 80.3% vs GPT-5.5 at 58.6%. That’s a roughly 22-point gap, with Fable ahead. SWE-Bench Pro is end-to-end resolution of real GitHub issues, so this is the closest thing we have to “can it actually close a real ticket across a real codebase.” Fable leads decisively.
SWE-Bench Pro: end-to-end GitHub issue resolution
Anthropic also reports Fable leading on legal reasoning (13.3% vs GPT-5.5’s 2.1%), spatial reasoning at 38.6%, state-of-the-art vision, and FrontierCode at roughly 2x Opus 4.8. So the verified picture is that Fable 5 is the stronger raw model on hard reasoning and codebase-wide coding.
Now let me be fair to GPT-5.5, because the SWE-Bench Pro number is not the whole story. GPT-5.5 is strongest at terminal-centric agentic coding inside Codex, and that’s exactly the work a lot of us do all day. Third-party comparisons report Terminal-Bench landing around Fable ~88% vs GPT-5.5 ~83% on a Codex CLI run, and OpenAI separately reported GPT-5.5 at ~82.7% on Terminal-Bench 2.0. Other figures floating around in third-party roundups — FrontierCode Diamond 29.3% vs 5.7%, OSWorld-Verified 85.0% vs 78.7%, GraphWalks @1M 68.1% vs 45.4%, GPT-5.5 hitting 85.0% on ARC-AGI-2 — are interesting but soft. I’d attribute them, not anchor on them.
My opinion, labeled as such: the durable read here is SWE-Bench Pro plus pricing. The exact third-party figures move week to week and shouldn’t drive your architecture. If you build a routing decision on a benchmark that gets re-run next month, you’re going to be re-architecting next month.
What does each one cost?
Here’s where it gets practical. All prices are per million tokens, verified from the official announcements.
- Claude Fable 5: $10 input / $50 output. Roughly 2x GPT-5.5, and about 2x Opus 4.8.
- GPT-5.5 (Standard): $5 input / $30 output. GPT-5.5 Pro is $30 input / $180 output.
- Reference middle — Claude Opus 4.8: $5 input / $25 output.
Output price per 1M tokens
Fable does have a ~90% prompt-caching discount, which lands cached input near ~$1/M — and that genuinely matters for agent loops that re-send the same context turn after turn. If you’re running long sessions with a fat, stable system prompt, that discount takes a real bite out of the bill.
But here’s the cost reasoning that actually decides things at volume: GPT-5.5 is roughly half Fable’s per-token price and it uses about 40% fewer output tokens than GPT-5.4 on the same Codex tasks. Cheaper per token and fewer tokens means GPT-5.5 is frequently the lower cost-per-successful-task option. Per-token price is the sticker; cost-per-task is what hits your invoice. I dig into that math more in how much it costs to add AI to your app.
Context windows and token efficiency
Both models reach ~1,000,000 tokens of context. Fable 5 takes text, image, and file input; GPT-5.5 hits 1M via the API. So at the headline level, context size is a tie.
One verified caveat worth knowing: GPT-5.5 is capped at 400,000 tokens inside Codex specifically. The full 1M is API-only. If you’re living in the Codex CLI on a huge monorepo, that 400K ceiling is a real constraint — you’ll feel it before you feel anything on the Fable side.
The token-efficiency edge is GPT-5.5’s quiet superpower. Using ~40% fewer output tokens than GPT-5.4 on the same tasks, while matching per-token latency, means it’s not just cheaper per token — it does the job with fewer tokens overall. That efficiency compounds with the lower price, and it’s the main reason I treat GPT-5.5 as the sensible default for everyday agent work rather than reaching for the premium tier by reflex.
Where each one genuinely wins
No fence-sitting here. Each model has a real lane.
Claude Fable 5 wins on codebase-wide resolution (that SWE-Bench Pro gap is the proof), on long-horizon autonomy where it has to stay coherent across many steps, and on multi-stage, document-heavy work. The concrete example Anthropic cites is Stripe running a 50-million-line Ruby migration in a single day — that’s sustained agentic delivery at a scale most tools fall apart at.
GPT-5.5 / Codex wins on terminal-centric agentic coding, cost-efficiency, and token-efficiency. It’s a strong, cost-efficient default for the work most of us actually do most days.
Where each one genuinely wins
Claude Fable 5
- Codebase-wide issue resolution (SWE-Bench Pro 80.3%)
- Long-horizon autonomy across many steps
- Multi-stage, document-heavy work
- Stripe: 50M-line Ruby migration in one day
GPT-5.5 / Codex
- Terminal-centric agentic coding
- Cost-efficient (about half Fable's per-token price)
- Token-lean (~40% fewer output tokens than 5.4)
- Strong everyday default
Opinion, labeled: most of my daily work is not the hard tail. It’s edits, fixes, refactors, plumbing. For that, the cost-efficient model wins on economics and rarely loses on quality. I reach for Fable when the task is genuinely long-horizon or spans the whole codebase — not on reflex. And since both have ~1M context, context size alone isn’t the tiebreaker. Task shape is.
Fable’s fine print: the caveats builders should know
Two verified things about Fable that belong in any honest comparison.
The safety fallback. Fable 5 auto-routes cybersecurity, biology, and distillation queries down to Claude Opus 4.8, and the user gets notified when it happens. So on those specific topics you effectively get Opus 4.8 — and you’re billed at Opus 4.8 rates ($5/$25) for that traffic, not Fable rates. The reassuring part: more than 95% of sessions never fall back. For most builders this is an edge case, not a daily tax. But if your product lives in security tooling, it’s a thing to design around.
Mandatory 30-day data retention. This one applies to all Fable/Mythos traffic — even customers who have zero-retention agreements in place. If you’re in a regulated or sensitive workload where zero-retention is a contractual or compliance requirement, this is the caveat to read twice. It’s not a dealbreaker for most teams, but it’s a real tradeoff and you should know it before you route sensitive data through Fable. Routing and cost-cap patterns like this are exactly what I cover in running LLMs in production.
FAQ: quick answers builders ask
Is Fable 5 worth 2x the price? Only for the hard long-horizon / codebase-wide tail where the SWE-Bench Pro gap actually pays off. For everything else, GPT-5.5 or Opus 4.8 is the cost-efficient pick.
Which is cheaper per task? Usually GPT-5.5 — roughly half the per-token price and about 40% fewer output tokens. Sticker price and token count both go its way.
Do they both have 1M context? Yes — but GPT-5.5 is limited to 400K inside Codex. The full 1M is API-only.
Does Fable ever hand off to another model? Yes. Cybersecurity / biology / distillation queries auto-route to Opus 4.8 (billed at Opus rates), and the user is notified. Over 95% of sessions never hit it.
What about data retention with Fable? A mandatory 30-day retention applies to all Fable/Mythos traffic, even under zero-retention agreements.
If you want the deeper breakdown of what Fable 5 actually is, I wrote that up in Claude Fable 5 explained.
The verdict: route by task, not by hype
There is no single winner here, and anyone telling you there is hasn’t run both at volume.
- Default to GPT-5.5 (or Opus 4.8) for cost-efficient everyday agentic and terminal work. This is most of your traffic.
- Reach for Claude Fable 5 on the hardest long-horizon and codebase-wide jobs — the ones where that 80.3% vs 58.6% gap clearly earns the 2x premium.
- Use GPT-5.5 Pro for accuracy-critical deep dives where you’d rather pay more than be wrong.
Route by task shape
Bottom line, and this is opinion: the win isn’t crowning a model. It’s knowing which task shape goes to which model and wiring your stack so the routing happens automatically. Fable for the hard tail, GPT-5.5 for the rest, Pro when being wrong is expensive. If you’re building agents and want more of this kind of practical engineering, I keep more AI engineering guides here.