What is Claude Opus 5?
Claude Opus 5 is Anthropic's July 2026 Opus model for agentic coding, deep reasoning, and enterprise work. Released July 24, 2026, it replaces Claude Opus 4.8 at the same per-token price and comes close to Claude Fable 5, Anthropic's top model, at half the price. On coding and knowledge work it goes further than close: it beats Fable 5 outright on Frontier-Bench v0.1 and ARC-AGI 3. It takes text and image input, outputs text, and carries a 1M token context window with up to 128k output tokens. On Unifically it runs as anthropic/claude-opus-5.
What's new in Claude Opus 5
Beats Claude Fable 5 on Frontier-Bench v0.1
Claude Opus 5 scores 43.3% against Claude Fable 5's 33.7% and Claude Opus 4.8's 18.7% on Anthropic's hardest agentic coding evaluation. More than double the previous Opus, at the same per-token price.
#1 on the Artificial Analysis Intelligence Index
The independent Artificial Analysis index scores Claude Opus 5 at 60.7, ahead of Claude Fable 5 at 59.9, GPT-5.6 Sol at 58.9, and Kimi K3 at 57.1, as of July 28, 2026. An independent cross-check of the intelligence-per-price claim.
30.2% on ARC-AGI 3
ARC-AGI 3 drops the model into interactive puzzles with no instructions and no stated goals. Claude Opus 5 scores 30.2% against 7.8% for the next-best model, the top result on the ARC Prize board as of July 24, 2026.
#1 on SWE-bench Verified
Vals AI measures 97.0% on SWE-bench Verified, the top spot on its board, ahead of GPT-5.6 Sol at 96.2%. Code review is strong too: it surfaces real bugs at a high rate per pass with few false positives, and stays accurate at low effort.
Thinking on by default
Every request thinks unless you turn it off, and the effort setting from low to max decides how deep. Extra effort converts into better results more reliably than on any earlier Opus, while low and medium keep strong quality at a fraction of the tokens.
1M token context, 128k output
The window holds a large repo, a long document set, or a full agent transcript at once. 1M tokens is both the default and the maximum, so there is no smaller variant to route around.
Best for
Multi-file feature work
Plans, edits, tests, and finishes end-to-end features without leaving stubs behind.
Code review and bug-finding
Real bugs at a high rate per pass, and still accurate at low effort for a cheap first pass.
Long-horizon agents
Stays on task across extended tool-use loops and checks its own work without being told.
Multi-agent systems
Runs subagent teams with writer-verifier patterns and few overwrite collisions.
Office and document work
Multi-sheet spreadsheets with working formulas and well-structured slide decks.
Chart and UI vision
Reads charts, documents, and diagrams, and rebuilds UI layouts from screenshots.
Use cases
Build a coding agent that takes a full feature spec, plans the change across many files, writes the tests, and reviews its own diff before handing it back. Run a review bot at low effort as a cheap first pass over every pull request, then a max effort pass on release branches. Put it behind an enterprise document pipeline that turns filings and spreadsheets into structured output with working formulas. It also fits computer-use agents: on OSWorld 2.0 it outperforms every other model at any given cost, passing Claude Fable 5's best result at about a third of the cost.
Limitations
Claude Opus 5 carries stronger cybersecurity safeguards than earlier Opus models. Requests that look like offensive security work can be declined, and some harmless security research gets caught. The API can re-run a declined request on a fallback model when you opt in.
Default responses and written files run longer than on Claude Opus 4.8. If your product needs short output, say so in the prompt; lowering effort does not reliably shorten the visible text.
Output is text only.
Claude Opus 5 vs Claude Fable 5
Claude Opus 5 wins Frontier-Bench v0.1 at 43.3% against 33.7%, and ARC-AGI 3 at 30.2% against 7.8% for the next-best model. On the Artificial Analysis Intelligence Index it edges Fable 5, 60.7 to 59.9, the top score on that board. At max effort it lands within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task, and on OSWorld 2.0 it passes Fable 5's best result at about a third of the cost. On Arena's Text board the picture is closer: Fable 5 holds #1 at 1508 Elo while claude-opus-5-max sits at #5 with 1495 on early votes, as of July 28, 2026. Fable 5 stays ahead on the hardest specialized evaluations and remains the top of the line. One practical difference: Fable 5 requires 30-day data retention, while Claude Opus 5 has no such requirement, which matters for strict-retention workloads.
When to use Claude Opus 5
Use Claude Opus 5 as the default Opus for coding, agents, and enterprise document work. It is the strongest model at its price, and it holds quality at low effort, so the same model can serve both the cheap pass and the careful one. Move up to Claude Fable 5 when the task sits at the edge of what models can do and cost is secondary. For short drafts and high-volume extraction, a smaller Claude model is still the cheaper fit.
API examples
Call Claude Opus 5 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Parameter docs live at docs.unifically.com/models/llm/overview.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "anthropic/claude-opus-5",
"messages": [
{ "role": "user", "content": "Review this diff and list every bug you find, with confidence and severity." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
anthropic/claude-opus-5, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
A 1M token context window, which is both the default and the maximum, with up to 128k output tokens per response. Instruction following and tool calling hold up across the whole window.
Yes. It takes text and image input in one request and reads charts, documents, diagrams, and screenshots. Output is text.
Yes, by a wide margin at the same per-token price. It scores 43.3% on Frontier-Bench v0.1 against 18.7% for Claude Opus 4.8, tops SWE-bench Verified at 97.0% on the independent Vals AI board, and improves every life-sciences evaluation Anthropic reports.
Thinking is on by default, and the model decides how much to think per request. Depth is set with the effort control, from low to max. Turning thinking off is only allowed at effort high or below.
On coding and knowledge work, mostly yes. It wins Frontier-Bench v0.1 at 43.3% against 33.7%, edges Fable 5 on the Artificial Analysis Intelligence Index at 60.7 to 59.9, and matches Fable 5's CursorBench 3.2 peak within 0.5% at half the cost per task. Fable 5 stays ahead on the hardest specialized evaluations.
Two things. Stronger cybersecurity safeguards can decline security-adjacent requests, so handle refusals and consider the built-in fallback option. And default responses run longer than on earlier Opus models, so prompt for short output where it matters.
