What is Claude Opus 4.7?
Claude Opus 4.7 is Anthropic's April 16, 2026 Opus model, built for agentic coding, computer use, knowledge work, and long-context reasoning. It sits between Opus 4.6 and the newer Opus 4.8 in the Opus line, and runs on Unifically as anthropic/claude-opus-4-7. It accepts text and image input and outputs text, with a 1M token context window and a 128k max output. The step up from Opus 4.6 is large: a near 7-point gain on SWE-bench Verified, the first high-resolution image input in the Claude line, and adaptive thinking that spends reasoning only where a turn needs it.
Key features of Claude Opus 4.7
Big coding gains
It scores 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro, up from 80.8% and 53.4% on Opus 4.6, and ahead of GPT-5.4 and Gemini 3.1 Pro on the Pro set.
High-resolution vision
Image input goes up to 2,576 pixels on the long edge, about 3.75 megapixels and over 3x the prior Claude limit. Small UI text, dense charts, and fine document detail stay readable.
Stronger computer use
It reaches 78.0% on OSWorld-Verified, up from 72.7% on Opus 4.6 and ahead of GPT-5.4 at 75.0%. Agents that click, type, and read screens get a clear accuracy bump.
Better knowledge work
It scores 64.4% on Finance Agent v1.1 and posts the top GDPval-AA result in its comparison set, with visible gains on office tasks like spreadsheet edits and document redlining.
1M token context window
Large codebases, long contracts, or many research papers fit in one context, with a callable max output of 128k tokens and better long-context reasoning than Opus 4.6.
Adaptive thinking
The model decides when and how deeply to reason on each turn, staying fast on simple lookups and spending more effort on hard multi-step problems.
Best for
Large-codebase agents
Multi-step fixes and refactors where the model must hold a full repo and follow through.
Computer-use agents
Browser and desktop-style tasks that read screens and click targets, with clear action limits.
Chart and document vision
Screenshots, scientific figures, and dense PDFs read at full resolution.
Finance and analysis work
Agent workflows over filings, models, and spreadsheets.
Long document review
Contracts, research packs, and multi-file jobs held in one 1M-token context.
Tool-heavy pipelines
Multi-tool orchestration, backed by a 77.3% MCP-Atlas score.
Use cases
Build a coding agent that traces a bug across a large repo, writes the fix, and verifies its own work over a long run. Use it for computer-use automation: form filling, internal tools, and desktop-style tasks where the higher image resolution makes small buttons and labels readable. It fits document-heavy work too, like contract review, financial research, and chart reading, where CharXiv scores of 82.1% without tools and 91.0% with tools show up as fewer misread figures. For product teams, Claude Opus 4.7 is a strong base for agents that plan multi-step jobs, call many tools in sequence, and report progress in plain language.
Limitations
Web research is a relative weak spot. Claude Opus 4.7 scores 79.3% on BrowseComp, down from 83.7% on Opus 4.6 and behind GPT-5.4 Pro at 89.3%. For multi-page web synthesis, test before you commit.
Built-in safeguards can refuse requests that look like high-risk cybersecurity work. Legitimate security tasks near that line may trip false positives.
The updated tokenizer counts roughly 1.0 to 1.35x more tokens than Opus 4.6 for the same text. Re-check token budgets and output caps instead of reusing old numbers.
High-resolution images cost more. A full-resolution image can use up to about 3x the image tokens of prior models, so downscale inputs when the extra detail is not needed.
Claude Opus 4.7 vs Claude Opus 4.6
Claude Opus 4.7 is a clear step up on coding, computer use, vision, and knowledge work. It scores 87.6% on SWE-bench Verified versus 80.8% for Opus 4.6, 64.3% on SWE-bench Pro versus 53.4%, and 78.0% on OSWorld-Verified versus 72.7%. Chart reading jumps most: 82.1% on CharXiv without tools versus 69.1%. Opus 4.6 keeps one win, web research, where it leads on BrowseComp. Pricing per token is unchanged between the two, though the new tokenizer means the same text costs somewhat more tokens on 4.7.
When to use Claude Opus 4.7
Use Claude Opus 4.7 when the job needs deep multi-step work: hard coding tasks, computer-use agents, vision over dense documents, or finance-grade analysis. Pick it over Opus 4.6 for anything agentic or vision-heavy. Move to Opus 4.8 when you want Anthropic's current Opus with better tool triggering and long-run coherence, or drop to a Sonnet model when speed and cost matter more than peak depth.
API examples
Call Claude Opus 4.7 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-opus-4-7.
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-4-7",
"messages": [
{ "role": "user", "content": "Trace this failing integration test across the repo and propose a fix." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
anthropic/claude-opus-4-7, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
Text and image input with text output. It reads images up to 2,576 pixels on the long edge, so screenshots, charts, and dense documents stay sharp instead of getting downscaled.
1M tokens, with a callable max output of 128k tokens. That fits large codebases, long filings, or a full research pack in one run.
Yes. It scores 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro, both large jumps over Opus 4.6, with better follow-through on long multi-step fixes.
Yes. It scores 78.0% on OSWorld-Verified, ahead of Opus 4.6 at 72.7%. The higher image resolution helps agents read small UI text and click the right targets. Real apps still have edge cases, so set clear action limits.
Yes. Claude Opus 4.7 uses adaptive thinking, so it decides when and how deeply to reason per turn instead of spending a fixed token budget on every request.
Web research is a relative weak spot, high-risk security requests can be refused by built-in safeguards, and the updated tokenizer counts more tokens than Opus 4.6 for the same text, so give output limits extra headroom.
