Released 2026-04-16 by Anthropic

Claude Opus 4.7
Coding / Vision / Agentic Three-Axis Upgrade

64.3% on SWE-bench Pro, 3.75 MP vision resolution, new xhigh effort tier, native 1M context — at the same price as Opus 4.6 ($5/$25 per MTok)

#Opus 4.7 #claude-opus-4-7 #agentic coding #high-res vision #long-horizon tasks #migration must-read

Key Highlights

+13%
+13% Coding

+13% over Opus 4.6 on a 93-task coding benchmark, including 4 tasks neither 4.6 nor Sonnet 4.6 could solve

64.3%
SWE-bench Pro 64.3%

6.6 points ahead of GPT-5.4 and 10.1 points ahead of Gemini 3.1 Pro on agentic coding

3.75 MP
3x Vision Resolution

Image long edge up from ~800px to 2576px (3.75 MP), a step-change for computer-use scenarios

$5 / $25
Same Price

$5 input / $25 output per MTok — identical to Opus 4.6, no price increase

Coding Leap: 28 Early Partners Validate

From long-horizon autonomy to complex tool calls, Opus 4.7 turns 'must watch closely' code work into 'hands-off'

GitHub: +13% on 93 Tasks

13% higher than Opus 4.6 on GitHub's internal 93-task coding benchmark, with 4 tasks neither 4.6 nor Sonnet 4.6 could solve

Cursor: 70% on CursorBench

Cursor's internal benchmark lifted from Opus 4.6's 58% to 70%

Notion: +14% Accuracy, 1/3 Tool Errors

Notion reports +14% accuracy, fewer tokens, tool-call errors reduced to 1/3; first model to pass Notion's implicit-need test

Cognition (Devin): Coherent Hours-Long Work

Opus 4.7 can work coherently for hours without giving up on hard problems

Rakuten: 3x Production Tasks Solved

On Rakuten-SWE-Bench, Opus 4.7 solves 3x the production tasks compared to Opus 4.6

⭐ Imbue: Autonomously Built a Rust TTS

Opus 4.7 built a full Rust TTS engine from scratch — neural net, SIMD kernels, browser demo — and used a speech recognizer to reverse-validate against the Python reference

Vision Capability Breakthrough

Image long edge up from ~800px to 2576px (3.75 MP), 3x over prior Claude models. Send images directly via API, no parameter switch needed

Computer-use Agents Reading Dense Screenshots

Higher resolution lets agents read more UI detail in a single view, reducing scroll/re-capture cycles

Complex Chart Data Extraction

Multi-layer nested charts, tables, dashboards — axis labels and fine details become readable

Document OCR & Layout Recognition

PDFs and scans with small text, footnotes, handwritten annotations — extract text and structure in one pass

UI Screenshot Pixel-Level Comparison

Design-to-implementation comparison, UI regression detection, and other pixel-precise scenarios

Newly Launched Capabilities

xhigh

New Effort Tier: xhigh

A new tier between high and max for finer-grained reasoning-depth vs latency trade-off. Claude Code defaults to xhigh

/ultrareview

/ultrareview Deep Code Review

New Claude Code command: an independent review session that runs through changes end-to-end, finding bugs and design issues

task_budgets

Task Budgets (API public beta)

Developers can set token budgets for Claude to self-allocate priorities on long tasks

auto_mode

Auto Mode Extended to Max Users

A classifier evaluates tool calls before execution — safe ones pass through, risky ones get intercepted for Claude to re-plan

Migration Guide (⭐ Key)

Upgrading from Opus 4.6 to Opus 4.7 is a drop-in replacement (change model ID to claude-opus-4-7), but 4 things to plan ahead

1. Tokenizer Switch

New tokenizer uses ~1.0-1.35x tokens for the same input (content-dependent). Re-evaluate token budgets, don't reuse old numbers

2. Stricter Instruction Following

Opus 4.7 executes instructions literally, no more 'charitable interpretation'. Old prompts may produce unexpected results; prompts and harnesses need re-tuning

3. Thinking API Migration

thinking={type:"enabled", budget_tokens:N} is deprecated; use thinking={type:"adaptive"} with effort parameter

4. Clean Up Old Beta Headers

effort-2025-11-24, fine-grained-tool-streaming-2025-05-14, interleaved-thinking-2025-05-14 are now GA — remove these beta headers

Code Example: Thinking API Migration
❌ Deprecated
client.messages.create(
    model="claude-opus-4-7",
    thinking={"type": "enabled", "budget_tokens": 10000}
)
✅ Recommended
client.messages.create(
    model="claude-opus-4-7",
    thinking={"type": "adaptive"},
    effort="xhigh"   # new in 4.7
)

vs GPT-5.4 / Gemini 3.1 Pro

Same-tier flagship comparison (based on Anthropic's published benchmarks)

Metric Opus 4.7 GPT-5.4 Gemini 3.1 Pro
SWE-bench Pro 64.3% 57.7% 54.2%
Input $ / MTok $5 See OpenAI See Google
Output $ / MTok $25 See OpenAI See Google
Context window Native 1M 272K / 1M beta 1M
Max output 128K tokens

Get Opus 4.7 via QCode.cc

Stable API proxy with official pricing, ready to use

Same Price $5/$25

QCode.cc bills at Anthropic's official rates with no multiplier markup

Full Support for New Parameters

Full pass-through of xhigh effort, adaptive thinking, and other new Opus 4.7 parameters

Drop-in Switch 4.6 to 4.7

Change model ID from claude-opus-4-6 to claude-opus-4-7, no other config changes needed

China-Direct with Failover

Multi-node smart routing + circuit breakers, avoiding instability of direct official API access from within China

Try Opus 4.7 Now

Sign up for QCode.cc to get a stable Claude Opus 4.7 API proxy