Claude Opus 5 Review — Benchmarks, Pricing, and What Changed From Opus 4.8
Anthropic shipped Claude Opus 5 on July 24, 2026 — API model ID claude-opus-5 — and it did something the last few Opus point releases did not: it took the outright top spot on the Artificial Analysis Intelligence Index at 61, one point ahead of the pricier Claude Fable 5 and two ahead of GPT-5.6 Sol. It did that at the same $5 / $25 price as Opus 4.8, the model it replaces. This is a review of what actually changed, which numbers to trust, and whether it is worth switching.
TL;DR verdict
| Claude Opus 5 | |
|---|---|
| Released | July 24, 2026 |
| Price | $5 / 1M input · $25 / 1M output (unchanged from Opus 4.8) |
| Context window | 1,000,000 tokens · 128k max output |
| Knowledge cutoff | May 2026 — the most current of any Claude model |
| Headline result | #1 on the Artificial Analysis Intelligence Index (61) |
| Coding | SWE-bench Pro 79.2% (vs Opus 4.8’s 69.2%) · SWE-bench Verified 96.0% |
| Availability | Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry; new default on Claude Max |
| Best for | Long-horizon agentic coding, agentic knowledge work, research |
| Skip if | You are happy on a cheaper mid-tier model and your work is not agentic |
If you do not read past this: Opus 5 is the clearest upgrade the Opus line has shipped in months, the price is unchanged, and it matches or beats the twice-as-expensive Fable 5 on most benchmarks — so making claude-opus-5 your default is close to a free win for agentic workloads.
What actually changed
Opus 5 is not a quiet point release like 4.7 → 4.8 was. Three things stand out: the top-of-leaderboard Intelligence Index result, a real jump on contamination-resistant coding benchmarks, and a price/performance story that undercuts Fable 5.
Benchmarks — a real coding and agentic step
Anthropic positioned Opus 5 as an agentic coding and knowledge-work model, and both the vendor numbers and the independent measurements back that up:
| Benchmark | Opus 5 | Opus 4.8 | Source | Notes |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 61 | — | Artificial Analysis | #1 of 187 models; ahead of Fable 5 (60) and GPT-5.6 Sol (59) |
| SWE-bench Pro (contamination-resistant) | 79.2% | 69.2% | Anthropic | +10 pt — the number to weight most |
| SWE-bench Verified (500 problems) | 96.0% | 88.6% | Anthropic | Near the top of the Verified set |
| Terminal-Bench 2.1 | 89.1% | 74.6% | Artificial Analysis | Adaptive Reasoning, max effort |
| Humanity’s Last Exam (no tools) | 56.3% | — | Artificial Analysis | 64.7% with tools per the system card |
The SWE-bench Pro number is the one to trust. The original SWE-bench Verified set is increasingly contaminated, so the ~10-point Pro gain over Opus 4.8 is a far better signal of real-world coding improvement than the Verified figure. The Terminal-Bench 2.1 jump — from the mid-70s to 89.1% — is the other result that matters for anyone running agents inside a shell, where recovering from a failed command is most of the job.
For the full cross-vendor picture — GPQA Diamond, MMMU, tau-bench, pricing, and context windows side by side — see the AI Models Leaderboard, where Opus 5 now sits at the top of the composite ranking. The benchmark cells on that page mark which figures are vendor-published, which are independently measured, and which are conservative estimates; Anthropic did not publish a standalone GPQA Diamond cell for Opus 5, so that column is flagged as an estimate and excluded from scoring.
It undercuts Fable 5
The strategic move in this release is pricing. Claude Fable 5 launched at $10 / $50 per million tokens as Anthropic’s premium tier. Opus 5 matches or beats Fable 5 across most published evaluations — it edges ahead on the Intelligence Index (61 vs 60) and is effectively level on Humanity’s Last Exam (56.3% vs Fable 5’s ~56.5% without tools) — while costing half as much. For most agentic and knowledge-work traffic, Opus 5 is now the more cost-effective frontier pick, and Fable 5 becomes a narrower choice for the specific workloads where it still leads.
Effort controls and extended thinking
Opus 5 keeps the effort model introduced with 4.8: extended thinking is on by default, and a per-request effort toggle (low / medium / high) lets you dial reasoning compute up for hard problems or down for routine tasks. Artificial Analysis reports the headline Intelligence Index at max effort, so expect the strongest results — and the highest token spend — at the top of that range. Output speed at max effort is modest (Artificial Analysis measured roughly 53 tokens/sec), which is the usual trade for heavy reasoning; drop the effort level for latency-sensitive work.
Pricing — what it costs
Pricing is unchanged from Opus 4.8, which — given the gains — is the quiet headline:
- Input: $5.00 per 1M tokens
- Output: $25.00 per 1M tokens
- Prompt caching: $6.25 write / $0.50 read per 1M tokens
- Context window: 1,000,000 tokens · max output: 128,000 tokens
At the same price as the model it replaces, with a measurable Intelligence Index and SWE-bench Pro lead, the cost story is straightforwardly positive. If your work is not agentic and you are price-sensitive, a cheaper model on the leaderboard — such as Claude Sonnet 5 — may still be the rational pick; Opus-class pricing only pays for itself when you are using the autonomy.
Who should upgrade
- Agentic coding teams: Yes. The SWE-bench Pro and Terminal-Bench 2.1 gains both target the exact failure modes that matter for autonomous, multi-file work. Switch your default to
claude-opus-5. - Agentic knowledge work / research: Yes. The top Intelligence Index score reflects broad gains across science, reasoning, and long-context retrieval.
- Current Fable 5 users: Re-evaluate. Unless you rely on a workload where Fable 5 specifically leads, Opus 5 gives you comparable intelligence at half the price.
- Chat-only users: Optional. There is no penalty to upgrading, but the general-conversation delta over Opus 4.8 is smaller than the agentic-coding delta.
- Anyone with hard-coded scaffolding: Re-test first. If your prompts pin a specific effort behavior or assume Opus 4.8’s output style, validate before moving production traffic.
How to switch
Point your client at the new model ID and decide whether to pin an effort level:
from anthropic import Anthropic
client = Anthropic()
resp = client.messages.create( model="claude-opus-5", # was claude-opus-4-8 max_tokens=4096, # extended thinking is on by default; effort defaults to high extra_body={"effort": "high"}, # "low" | "medium" | "high" messages=[{"role": "user", "content": "Refactor this module and run the tests."}],)print(resp.content[0].text)In Claude Code, Opus 5 is the new default on Claude Max and the effort menu exposes the same levels. See Agent Instructions for how to scope a CLAUDE.md so the upgraded model stays inside your conventions.
FAQ
How much does Opus 5 cost? $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8 — with a 1M-token context window and 128k max output.
Is Opus 5 better than Fable 5? On the Artificial Analysis Intelligence Index it scores 61 to Fable 5’s 60, and it matches or beats Fable 5 on most published benchmarks while costing half as much. Fable 5 remains a distinct premium tier for the workloads where it still leads.
How much better is it at coding than Opus 4.8? Anthropic reports 79.2% on SWE-bench Pro (up ~10 points from 69.2%) and 96.0% on SWE-bench Verified; Artificial Analysis measures 89.1% on Terminal-Bench 2.1 at max effort.
Should I upgrade from Opus 4.8? For coding and agentic work: yes — same price, real gains. For chat: optional. Re-test any scaffolding that hard-codes effort or depends on 4.8’s output style first.
How does it compare to GPT-5.6 Sol? Opus 5 leads on the composite Intelligence Index (61 vs 59). The full cross-vendor table is on the leaderboard.
Continue reading
- AI Models Leaderboard — Opus 5 versus 70+ models on benchmarks, pricing, and context window.
- Claude Opus 4.8 Review — the model Opus 5 replaces, and the effort controls it inherits.
- Claude Fable 5 Review — the premium tier Opus 5 now matches at half the price.
- Claude Sonnet 5 Review — the cheaper mid-tier sibling for high-volume agent traffic.
- GPT-5.6 Sol, Terra & Luna Review — OpenAI’s frontier line that Opus 5 edges on the Intelligence Index.
- LLM Benchmark Comparison 2026 — how to read SWE-bench Pro, Terminal-Bench, and HLE without getting fooled.
- All Reviews — index of every head-to-head review on the site.