Claude Opus 5 Benchmarks: What the Numbers Actually Tell You
Anthropic has published benchmark numbers for Claude Opus 5, compared against its own Fable 5 and Opus 4.8 and against OpenAI's GPT-5.6 Sol. We run AI agents in production every day — our own operations included — so we read these tables differently than a launch-day headline: not "who won", but which capabilities moved, and what that changes for teams deploying this stuff. Here is the full table, then the reading.
The numbers (as published by Anthropic)
| Benchmark | Opus 5 | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| Agentic terminal coding (Frontier-Bench v0.1) | 43.3% | 33.7% | 21.1% | 34.4% |
| Knowledge work (GDPval-AA v2, Elo) | 1861 | 1747 | 1593 | 1736 |
| Novel problem-solving (ARC-AGI-3) | 30.2% | — | 1.5% | 7.8% |
| Agentic search (BrowseComp) | 90.8% | 87.4% | 84.3% | 90.4% |
| Humanity's Last Exam — no tools | 56.3% | 56.5% | 49.8% | — |
| Humanity's Last Exam — with tools | 64.7% | 63.9% | 57.9% | — |
| Computer use (OSWorld 2.0) | 70.6% | 66.1% | 55.7% | 62.6% |
| Agentic coding (DeepSWE v1.1) | 68.8% | 69.7% | 59.0% | 72.7% |
| Agentic coding (FrontierCode v1.1, Main) | 53.4% | 53.5% | 46.5% | 47.5% |
| Business workflows (AutomationBench) | 26.0% | 17.4% | 17.0% | 18.1% |
| Legal (Legal Agent Benchmark, held-out) | 11.7% | 13.3% | 10.4% | 2.5% |
| Health (HealthBench Professional) | 59.8% | 66.0% (Mythos 5) | 57.4% | 60.5% |
| Biology (BioMysteryBench, hard) | 49.4% | 46.5% | 42.4% | — |
| Biology (BioMysteryBench, human-solved) | 90.1% | 89.0% (Mythos 5) | 88.5% | — |
Source: Anthropic's published comparison. Standard caveat applies: these are vendor-published numbers on vendor-selected benchmarks — treat them as directional until independent evaluations land. A note on names: Fable 5 and Mythos 5 share the same underlying model; Fable is the generally available variant with additional safety measures, Mythos the variant for approved organizations.
The three numbers that actually matter
- ARC-AGI-3: 1.5% → 30.2% in one generation. This is the outlier in the table. ARC-AGI is deliberately designed to resist memorization — it tests solving genuinely novel puzzles. A 20x jump (with GPT-5.6 at 7.8%) is the kind of discontinuity that usually signals a real capability change, not benchmark tuning. If it holds up independently, "the model falls apart on problems it hasn't seen" — the classic enterprise objection — just got materially weaker.
- AutomationBench: 26.0% vs everyone else's ~17-18%. Business-workflow automation is the benchmark closest to what our clients actually buy AI for — multi-step processes across real tools. Two readings: Opus 5 pulled clearly ahead, and the absolute number is 26%. Three quarters of benchmark-grade business workflows still fail. This is exactly why we design agents with narrow scopes, checkpoints, and human escalation instead of "automate the department".
- OSWorld 2.0: 70.6%. Computer use crossed from "demo" territory (55.7% one generation ago) toward "usable with supervision". Screen-driving agents for legacy systems without APIs are becoming a real integration option — we already prototype these for internal workflows.
Where the honest asterisks go
- GPT-5.6 Sol wins DeepSWE (72.7% vs 68.8%). On one of the two agentic-coding benchmarks, OpenAI leads — while Opus 5 takes the other (FrontierCode) and terminal coding by a wide margin. Read: coding leadership is contested and benchmark-dependent, which is precisely why multi-model routing beats single-vendor loyalty.
- Fable 5 matches or beats Opus 5 in spots — HLE without tools, FrontierCode, legal. The mid-tier model at (typically) lower cost being this close matters more for your bill than the flagship's records: most production workloads don't need the flagship.
- Legal at 11-13% across the board is the table's quiet warning. On held-out legal agent tasks, every model fails ~7 times out of 8. Anyone selling you an autonomous legal agent in 2026 is selling ahead of the data.
What this changes in practice
Our operational takeaways, from running AI agents in production (our own AiMon and agent layer included): (1) Route by task, not by brand. The table itself argues for it — different models win different rows, and mid-tier models are close enough that sending everything to a flagship is a billing decision, not an engineering one. Our which-AI-for-what guide covers the routing logic. (2) Re-run your evals, not theirs. Vendor benchmarks say a new model exists; only your regression suite on your actual prompts says you should switch. Teams without an eval set are choosing models by press release. (3) Capability is outpacing operations. The gap between a 30% ARC score and a reliable production system is monitoring, guardrails, cost control and fallbacks — the unglamorous layer where AI projects actually live or die. That layer is our day job; the model layer improving just raises what it can carry.
Frequently asked questions
Is Opus 5 better than GPT-5.6?
On Anthropic's published table, Opus 5 leads most rows — decisively on novel problem-solving, computer use, and business workflows — while GPT-5.6 Sol wins one agentic-coding benchmark (DeepSWE). Contested, benchmark-dependent, and worth verifying on your own workloads.
What is the difference between Fable 5 and Mythos 5?
They share the same underlying model. Fable 5 is generally available and includes additional safety measures for dual-use capabilities; Mythos 5 is available without those measures to approved organizations only.
Why is the ARC-AGI-3 result significant?
ARC-AGI is designed to resist memorization and test genuinely novel problem-solving. Jumping from 1.5% (Opus 4.8) to 30.2% in one generation — versus GPT-5.6's 7.8% — suggests a real capability shift rather than benchmark tuning, pending independent verification.
Should we switch models in production?
Not on a press release. Re-run your own evaluation suite on your actual prompts and tools, compare cost per task (not per token), and check the mid-tier first — the published numbers show mid-tier models close to flagships on many tasks.
Bottom line
Opus 5's table shows real movement — novel problem-solving and workflow automation most of all — with honest asterisks: a contested coding crown, mid-tier models nipping at flagship heels, and legal agents still failing 7 of 8 tasks. The models keep getting better at their job; whether they get better at yours is decided by evals, routing, and the operational layer around them.
Running — or planning to run — AI workloads in production? We operate our own AI agents daily and run Bedrock and LLM operations for clients as part of AWS Managed Services: model routing, cost control, evals, and the monitoring layer that keeps it honest. Start with a conversation.