We ran Claude Code with Anthropic's Opus 5.5, released on September 22, on the same real-world coding tasks we use for the Agent Security League. It scores 68.7% on generating functional code (FuncPass) and 33.5% on generating functional and secure code (SecPass). This is a hair above its predecessor Opus 5 on security (32.4%) but below it on functionality (73.7%), and behind Fable 5.1 on both axes (87.2% / 37.4%).
While Anthropic claims Opus 5.5 achieves a new state-of-the-art for coding and outpaces Fable in many benchmarks, we did not find that held true for our benchmark. On the full board Opus 5.5 ranks third on generating secure code, with Claude Code running Fable 5.1 at the top and OpenAI Codex with GPT-6 Astra (34.6%) sitting in between. We also found that Opus 5.5 only places nineteenth on generating code that passes functional tests, which is a fairly mediocre showing.
What it does better than anything we have run is efficiency, which roughly matches the marketing claims for the model: about a quarter of Opus 5's time to complete tasks, roughly a tenth of its cost, and not a single timeout. The catch is that the anti-cheating pipeline confirmed 51 training-recall cheats, the highest count on our board; had those memorized solves counted, Opus 5.5 would lead the security ranking by a wide margin.
This post checks Anthropic's launch claims against what we measured, compares Opus 5.5 with Opus 5 and Fable 5.1 on score, speed, tokens and cost, and walks through where the memorization shows up.
Key takeaways
- Third on security, nineteenth on functionality. Claude Code + Opus 5.5 reaches 68.7% FuncPass and 33.5% SecPass — +1.1 SecPass points over Opus 5 but −5.0 on FuncPass, and behind Fable 5.1 on both axes (87.2% / 37.4%).
- The largest memorization tax on the board. Had memorized solves counted, Opus 5.5 would score 94.4% / 52.5% and lead the security ranking by more than 11 points; removing them costs 25.7 FuncPass points and 19.0 SecPass points, versus 6.7 / 3.9 for Fable 5.1 and 18.4 / 14.0 for Opus 5. All 51 confirmed cheats are training recall; 37 of them were secure passes.
- Ten times cheaper than Opus 5, six times cheaper than Fable 5.1. Recorded prediction spend was $116 ($0.33 median per task) against ~$1,116 for Opus 5 and $672 for Fable 5.1 — the third-cheapest full run on our board — driven both by the lower list price and by 3× fewer output tokens per task.
- Fastest run we have recorded, with no timeouts. Mean 4.0 min / median 2.2 min per task; 79% of tasks finish under five minutes and none exceeded 31 minutes. Fable 5.1 averaged 9.5 min, Opus 5 14.1 min with 15 timeouts.
- A third of the tool calls. ~17 tool calls and ~33 assistant turns per task, against 42 / 73 for Fable 5.1 and 75 / 116 for Opus 5. Memorized solves are the quickest of all: 13.5 tool calls and 3.0 minutes on average.
- Launch claims, checked. "Cheaper" and "faster" hold up strongly; "new state-of-the-art" and "outpaces Fable" do not on secure coding once recall is discounted; the cyber safeguard was monitored and never triggered.
- Two sole security solves, one survives. Opus 5.5 alone securely fixed two tasks; one was blocked as memorization after the CVE identifier surfaced in the trajectory, the other stands as a legitimate, recall-seeded pass.
Introduction
Anthropic released Opus 5.5 on September 22, 2026 as a cheaper, faster Opus that, in its own benchmarks, outpaces the larger Fable model. It arrives two months after Opus 5 and three weeks after Fable 5.1 took the top of our board. That gives us a clean three-way comparison inside one harness: Claude Code with Opus 5, Fable 5.1 and Opus 5.5 on the same coding tasks, with the model as the only meaningful variable.
The efficiency story is real and larger than the launch post suggests: Opus 5.5 is not just cheaper per token, it also reads less, writes less and stops earlier, so a full run cost us $116 where Opus 5 cost over $1,100. In total it is 6x cheaper to run.
The capability story is more complicated. Before the anti-cheating pipeline has its say, Opus 5.5 passes as many functional tests as any combo we have run and more security tests than all of them, but a large share of those passes come from the model recognising the task. Our pipeline confirmed 51 instances where the fix was recalled rather than derived — and, prompted by two of them, we added a provenance check for CVE identifiers to the pipeline itself.
This post takes Anthropic's launch claims one by one, then compares the three Anthropic models on scores, speed, token consumption and cost, before turning to the cheating picture and the two tasks only Opus 5.5 solved securely.
Benchmark recap
Skip if you know this already.
We measure combos including a harness (Claude Code, Cursor, Codex, …) plus a frontier model on coding tasks inside real, complex projects. Each task involves code that was historically part of a security fix, but the combo is never told this; it is only asked to follow security best practices.
Each combo runs once per task, and we apply its patch in an isolated Docker environment. FuncPass means the patch passes the functional tests. SecPass means it also passes the hidden security tests introduced by the original vulnerability fix, so a secure result must first be functionally correct.
We also require the fix to come from the combo's own reasoning: recovering the known fix from git history, the web, a workspace copy, or training recall is treated as cheating. A multi-signal pipeline flags suspicious instances, and an LLM adjudicates each one. Confirmed cheating is removed from the score.
The benchmark also contains a small set of overly strict instances: their security tests demand implementation details that are exceptionally unlikely to be guessed independently. Rather than discard those instances entirely, we exclude them from the leaderboard denominator — they are not a fair measure of general secure coding ability — but keep them as traps. A pass is a tripwire for deeper inspection rather than automatic credit, and the agent’s trajectory can reveal a recall pattern that ordinary final-patch comparison misses. That is what happened in this run. For recall-then-diverge, we now also grade the edit trajectory — first/peak security overlap, suddenness, and string chronology — so early recall remains visible even after the final patch has diverged.
Opus 5.5 benchmark results
Opus 5.5 lands third on SecPass and nineteenth on FuncPass among the combos on the board. The more revealing comparison is with its Anthropic siblings, which all ran under the Claude Code harness on the same coding tasks, making the model the main variable. (The harness version is not identical: Opus 5.5 ran on Claude Code 2.1.280, Fable 5.1 on 2.1.258, Opus 5 on 2.1.156.)
Comparison with Fable 5.1 and Opus 5
Opus 5.5 edges past Opus 5 on SecPass (+1.1 points) but falls behind it on FuncPass (−5.0 points), and it trails Fable 5.1 by 18.5 FuncPass points and 3.9 SecPass points. The FuncPass deficit is almost entirely the cheating deduction: 46 of the 51 confirmed cheats fall inside the 179-task denominator, and every one of them had passed the functional tests.
Opus 5.5 also enters the security hall of fame with one entry: of the two tasks it alone fixed securely, the jupyter_core pass was disqualified as memorized because the model gave itself up by naming the vulnerability's CVE identifier before its first edit, while the rdiffweb rate-limiter rebuild stands as a legitimate solve that no other combo has managed.
Operationally the run was spotless: every prediction completed, zero timeouts, zero failed runs, zero API errors, and the model stayed claude-opus-5-5
throughout, no fallbacks to a smaller model. We also monitored Anthropic's cyber safeguard, which the launch post says applies to Opus 5.5 as it does to Fable; it did not trigger on any task.
With and without memorization
As memorization becomes more visible across frontier models, we present scores in two forms on the same denominator:
- Memorization kept: confirmed memorized solves still count as passes.
- Memorization removed: confirmed memorized solves are treated as cheating and do not count.
The numbers published on our leaderboard, and everywhere else in this post, are those with memorization removed.
With memorization kept, Opus 5.5 would be first on SecPass by more than 11 points and half a point ahead of Fable 5.1 on FuncPass. With it removed, it drops to third. No combo on our board loses as many points to recall as Opus 5.5 does; the previous high-water marks were Opus 5 itself (−18.4 / −14.0) and Fable 5.0 under Cursor (42 confirmed cheats, −17.9 / −14.0).
Anthropic's launch claims, checked against our results
The TechCrunch launch coverage makes a handful of concrete claims. Here’s what we observed:
- "A new state-of-the-art in coding and knowledge work performance." Not on secure coding: 68.7% / 33.5% sits behind Fable 5.1 on both axes and behind Opus 5 on FuncPass, and the only version of the score that would make it #1 is the one that counts recalled solves.
- "Outpaces the larger Fable model in many benchmarks." Not on ours: Fable 5.1 leads by 18.5 FuncPass points and 3.9 SecPass points; the only axis where Opus 5.5 is ahead is SecPass with memorization kept (52.5% vs 41.3%), and that surplus is memorization.
- "Significantly cheaper — $20 per million output tokens versus $25 for the previous model." Backed, and understated based on our analysis: the run cost $116 against ~$1,116 for Opus 5 and $672 for Fable 5.1, because the lower price compounds with far lower consumption — a third of Fable 5.1's output tokens per task and less than a quarter of Opus 5's.
- "Faster to run, reflecting a drop in the compute required to serve it." Backed: mean 4.0 minutes per task versus 14.1 for Opus 5 and 9.5 for Fable 5.1, with 79% of tasks done in under five minutes and zero timeouts.
Faster than anything we have measured, and no timeouts
While we don’t officially measure timeouts and speed for the agent security league, we did observe that Opus 5.5 is about 2.4x faster than Fable 5.1 and 3.5x faster than Opus 5 on the mean, and the distribution is compressed rather than merely shifted: 157 tasks finished in under five minutes (Fable 5.1: 74; Opus 5: 53), 179 in under ten, and the slowest took 30.2 minutes. The whole run took 5.8 hours of wall-clock time on six parallel workers.
Speed correlates with recall, though. The 51 memorized solves averaged 3.0 minutes and 13.5 tool calls; the other 149 tasks averaged 4.4 minutes and 17.6 tool calls. The model is quickest on the tasks it already knows.
Token consumption and where the cost goes
Claude Code reports cumulative cost and token use per session; we take the final report of each distinct session and sum sessions per task, the same method as in the Fable 5.1 post.
Per task (median): Opus 5.5 emitted 6.4k output tokens and read 0.37M cached tokens for $0.33; Fable 5.1 23.0k / 0.96M / $2.33; Opus 5 38.1k / 3.28M / $3.31.
The 5.8x cost gap to Fable 5.1 decomposes roughly in half. Anthropic's price cut accounts for part of it, $20 per million output tokens against Fable's $50, and the rest is consumption: Opus 5.5 generates about a third of the output tokens and touches under half the context of Fable 5.1 per task. Against Opus 5 the gap is close to 10×, with cache reads down almost 9×. For a full pass over the benchmark, that is the difference between a four-figure and a three-figure bill.
Tool use: a third of the calls
Opus 5.5 spends a flat effort regardless of outcome, 15.9 tool calls on secure passes and 17.1 on failures, the same pattern Fable 5.1 and Opus 5 showed — but at less than half Fable 5.1's volume and less than a quarter of Opus 5's. The low count is the efficiency story and the memorization story at once: a model that recognises the task does not need to read much before it writes the fix.
Conclusion
Opus 5.5 is the most cost-efficient Anthropic model we have run, by a wide margin. Only Composer 2.5 and Google Gemini 3.5 Flash under Cursor have cost less for a full pass over the benchmark, and both score far lower, and on that dimension the launch claims are conservative. On secure coding it is a strong Opus 5 rather than a new leader, and its memorization tax is the largest we have recorded. The two are connected: a model that recognises the task is both fast and, on this benchmark, not solving it.
What's next?
When you're ready to take the next step in securing your software supply chain, here are 3 ways Endor Labs can help: