Anthropic

Claude Opus 4.7

Provider

Anthropic

Status

Superseded

Context

1,000,000 tok

SWE-bench

87.6%

Price

$5 / $25 /MTok

Knowledge

2026-01

Claude Opus 4.7 was Anthropic’s flagship from 16 April 2026 until Opus 4.8 replaced it on 28 May 2026. It remains generally available and is still an excellent model — but if you’re choosing today, Opus 4.8 is the newer, marginally stronger option at the same price. This page covers Opus 4.7 as the direct predecessor: what it changed, where it still holds up, and what to know before deploying or migrating off it.

Opus 4.7’s headline was coding. SWE-bench Verified jumped from Opus 4.6’s 80.8% to 87.6%, and the harder, multi-language SWE-bench Pro leapt from 53.4% to 64.3% — clear of GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) (Anthropic, Vellum). Two other things made 4.7 notable beyond the numbers: it was the first broadly-released Claude to ship with automatic cybersecurity safeguards (the start of Anthropic’s Project Glasswing rollout), and it brought a large vision upgrade — 3.75-megapixel image support, more than triple prior Claude models.

Quick specs

TRY CLAUDE →

What’s new in Claude Opus 4.7

Opus 4.7 was a focused upgrade to Opus 4.6, not a broad sweep. Anthropic was explicit that its restricted Claude Mythos Preview remained more broadly capable; 4.7 targeted the things that break production agents (Anthropic).

Harder coding, handed off with confidence

The core pitch was that you could hand 4.7 the kind of hard, long-running coding work that previously needed close supervision. Anthropic highlighted rigorous handling of complex tasks, precise instruction-following, and a habit of verifying its own outputs before reporting back. Partners corroborated this: Cursor reported a jump from 58% to 70% on CursorBench, and several reported double-digit lifts in task resolution with fewer tool errors (Anthropic).

First model with automatic cyber safeguards (Project Glasswing)

Opus 4.7 was the first broadly-released Claude to ship with the Project Glasswing safeguard stack — automated systems that detect and block requests indicating prohibited or high-risk cybersecurity uses. Anthropic says it even experimented during training with differentially reducing offensive cyber capabilities relative to raw model scale. Legitimate security researchers (vulnerability research, pen-testing, red-teaming) can apply to the Cyber Verification Program. This was the test bed for safeguards Anthropic intends to carry to its more powerful Mythos-class models.

A big vision upgrade

Opus 4.7 accepts images up to 2,576 pixels on the long edge (~3.75 megapixels) — more than three times the resolution of prior Claude models. The effect on visual reasoning was the single largest benchmark jump in the table: CharXiv (scientific figure interpretation) rose from 69.1% to 82.1% without tools. One early partner (XBOW) saw visual acuity on its internal benchmark go from 54.5% to 98.5% (Anthropic). Anthropic also called 4.7 more “tasteful” on professional outputs — interfaces, slides and docs.

Effort control, memory and Claude Code tooling

Opus 4.7 introduced a new xhigh effort level between high and max for finer reasoning/latency control; Claude Code raised its default to xhigh. The model is also better at file-system-based memory across long, multi-session work. Alongside the launch, Anthropic shipped task budgets (public-beta API control over token spend on long runs), the /ultrareview Claude Code command for dedicated review sessions, and extended auto mode to Max users.

A tokenizer change worth planning for

Two migration caveats affect token usage. First, Opus 4.7 uses an updated tokenizer, so the same input maps to roughly 1.0–1.35x more tokens than Opus 4.6 (up to ~35% more, depending on content). Second, it thinks more at higher effort levels. Per-token price is unchanged, but real spend per task can rise — Anthropic recommends measuring on real traffic and provides a migration guide. Improved instruction-following also means 4.7 takes prompts more literally, so prompts tuned for older models may need re-tuning.

The Opus 4.7 model family

Like the rest of the modern Opus line, Opus 4.7 is a single API model (claude-opus-4-7) whose behaviour is shaped by two dials, not separate model IDs.

Within the wider Claude lineup, Opus 4.7 sits below the newer Opus 4.8 and the Mythos-class Fable 5, and above mid-tier Sonnet 4.6 and fast/cheap Haiku 4.5. At launch, Anthropic’s own Claude Mythos Preview was the more capable (but restricted) model above it.

Benchmark performance

Anthropic compared Opus 4.7 against Opus 4.6, GPT-5.4, Gemini 3.1 Pro and its restricted Claude Mythos Preview in the System Card. All figures below are Anthropic-reported.

Coding

The SWE-bench gains are the story: +6.8 points on Verified and +10.9 on the harder Pro variant, the largest coding jump of that generation and clear of every then-available competitor. Anthropic’s memorisation screens flagged a subset of SWE-bench problems; it says 4.7’s margin over 4.6 holds even with those excluded (Anthropic footnotes). *GPT-5.4’s 75.1% on Terminal-Bench is self-reported on a different harness and not directly comparable; on the common Terminus-2 harness (thinking disabled), Opus 4.7 edged Gemini 3.1 Pro.

Agentic, tool use and computer use

MCP-Atlas was Opus 4.7’s strongest competitive result — best-in-class tool use, the number that matters most for multi-tool orchestration agents. OSWorld-Verified rose 5.3 points and, paired with the vision upgrade, made 4.7 a meaningfully better computer-use model (within 1.6 points of Mythos Preview’s 79.6%). Finance Agent was state-of-the-art at release. The one clear regression: BrowseComp dropped 4.4 points versus Opus 4.6 — if your workload is research- and web-browsing-heavy, that’s worth knowing, and GPT-5.4 Pro or Gemini 3.1 Pro were the stronger picks there.

Reasoning, knowledge and vision

GPQA Diamond is saturated — the top four models sit in a 94.2–94.6% cluster, so 4.7’s small dip versus GPT-5.4 Pro and Gemini is noise. On Humanity’s Last Exam, 4.7 improved on 4.6 in both configurations and beat Gemini 3.1 Pro with tools, but trailed GPT-5.4 Pro and (clearly) Mythos Preview. CharXiv’s 13-point jump maps directly to the higher-resolution vision. Gemini led multilingual QA.

Independent verification

Artificial Analysis lists Opus 4.7 and independently runs GDPval-AA, where Anthropic claims state-of-the-art knowledge-work performance. As with most fresh frontier releases, the headline System Card figures are vendor-run; the most-cited independent explainer (Vellum) reproduces those numbers rather than re-running them.

Pricing breakdown

Opus 4.7 kept the standard Opus price — $5 input, $25 output per million tokens — and that same price carried forward to Opus 4.8 (Anthropic).

The catch isn’t the sticker price — it’s the tokenizer. Because Opus 4.7 maps the same input to up to ~35% more tokens than Opus 4.6, and reasons more at higher effort, the effective cost per task can rise even though the per-token rate is flat. Anthropic frames the net effect as favourable on its internal coding eval (token usage improved at each effort level), but advises measuring on real traffic. There is no long-context surcharge for the 1M window.

Cost comparison with contemporaries

How to access Claude Opus 4.7

Via API

Opus 4.7 is generally available as claude-opus-4-7 on the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(message.content)

Note the cyber-use safeguards: requests indicating prohibited or high-risk cybersecurity activity are automatically blocked. Legitimate security work can be unblocked via the Cyber Verification Program.

Via the Claude apps

Opus-tier access requires a paid plan; the free tier serves Sonnet 4.6 / Haiku 4.5 (Anthropic pricing).

On the apps you’ll now typically be defaulted to the newer Opus 4.8; Opus 4.7 remains available primarily via the API for teams that haven’t migrated.

How Claude Opus 4.7 compares

vs Claude Opus 4.8 (its successor)

Opus 4.8 (28 May 2026) is the direct upgrade, at the same $5/$25 price. It edges 4.7 on most benchmarks — e.g. SWE-bench Pro 69.2% vs 64.3%, SWE-bench Verified 88.6% vs 87.6% — and adds a notable honesty improvement (around 4x less likely to let its own code flaws pass unflagged), effort control on all plans, dynamic workflows in Claude Code, and a fast mode cut to $10/$50. The gains are real but, by Anthropic’s own framing, “modest.” For new work, prefer Opus 4.8. Opus 4.7 still makes sense if you have prompts/harnesses tuned to it and don’t want to re-validate, or pipelines pinned to claude-opus-4-7.

vs GPT-5.4

On Anthropic’s benchmarks, Opus 4.7 led GPT-5.4 on coding (SWE-bench Pro 64.3% vs 57.7%) and tool use (MCP-Atlas 77.3% vs 68.1%). GPT-5.4 Pro led research-style and hard-reasoning tasks: BrowseComp (89.3% vs 79.3%) and Humanity’s Last Exam with tools (58.7% vs 54.7%). Choose Opus 4.7 for coding and multi-tool agents; GPT-5.4 for research-heavy browsing workflows.

vs Gemini 3.1 Pro

Opus 4.7 led Gemini 3.1 Pro on coding (SWE-bench Pro 64.3% vs 54.2%; Verified 87.6% vs 80.6%), tool use (MCP-Atlas 77.3% vs 73.9%), finance (64.4% vs 59.7%) and HLE with tools (54.7% vs 51.4%). Gemini led multilingual QA (MMMLU 92.6% vs 91.5%) and BrowseComp (85.9% vs 79.3%). For the current Google flagship, see the Google hub — Gemini 3.5 Pro was announced after Opus 4.7’s comparison window and, as of early July 2026, remains in limited preview (general availability expected July).

The practical consensus

At release, Opus 4.7 was the strongest generally available coding and tool-use model — best-in-class MCP-Atlas, a big SWE-bench Pro jump, and a meaningful computer-use/vision upgrade — with one watch-out (BrowseComp regression) and a tokenizer that nudges costs up. Anthropic’s restricted Mythos Preview was more capable across the board. Today, Opus 4.8 occupies the flagship slot. See our best AI models and best AI for coding rankings for current standings.

Known limitations

Superseded by Opus 4.8. As of late May 2026, 4.7 is no longer Anthropic’s flagship. It’s still available and capable, but 4.8 is better at the same price.

BrowseComp regression. Agentic web research dropped from 83.7% (4.6) to 79.3% — a real step back for research-heavy agents; GPT-5.4 Pro and Gemini 3.1 Pro were stronger there.

Tokenizer raises token counts. Same input maps to up to ~35% more tokens than Opus 4.6, and higher effort means more thinking tokens. Per-token price is flat, but per-task spend can rise.

Prompt re-tuning needed. More literal instruction-following means prompts tuned for older models can behave unexpectedly.

Vendor-reported benchmarks. Headline figures are Anthropic-run via the System Card; independent composite scores were limited at the time of writing.

Cyber-use restrictions. Safeguards auto-block prohibited/high-risk security requests; legitimate researchers must go through the Cyber Verification Program.

Free tier excluded. Opus tier requires a paid plan.

Community reception

Reception from early-access partners was strongly positive, concentrated on agentic reliability and coding (Anthropic):

• Cursor (Michael Truell): “a meaningful jump… clearing 70% [on CursorBench] versus Opus 4.6 at 58%.”
• GitHub (Mario Rodriguez): +13% resolution on a 93-task coding benchmark, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve.
• Cognition / Devin (Scott Wu): “works coherently for hours, pushes through hard problems rather than giving up.”
• Notion (Sarah Sachs): “+14% over Opus 4.6 at fewer tokens and a third of the tool errors… the first model to pass our implicit-need tests.”
• XBOW (Oege de Moor): visual acuity “98.5%… versus 54.5% for Opus 4.6 — our single biggest Opus pain point effectively disappeared.”
• Vercel (Joe Haddad): “noticeably more honest about its own limits… does proofs on systems code before starting work.”

The recurring themes — fewer tool errors, end-to-end follow-through, loop resistance, and a more “opinionated” model that pushes back — line up with Anthropic’s framing of 4.7 as the Opus reliable enough to hand the hardest work to.

Version history

Anthropic positioned 4.7 explicitly as the bridge model where real-world cyber safeguards get tested ahead of an eventual broad release of Mythos-class models.

FAQ

Is Claude Opus 4.7 still worth using in June 2026?

It’s still strong, but it’s been superseded by Opus 4.8, which is marginally better at the same $5/$25 price. Use 4.7 if you have prompts/harnesses tuned to it or pipelines pinned to claude-opus-4-7; otherwise prefer 4.8 for new work.

How much does Claude Opus 4.7 cost?

$5 per million input tokens and $25 per million output — unchanged from Opus 4.6. Note the updated tokenizer can map the same input to ~1.0–1.35x more tokens, so per-task spend may rise even at a flat per-token rate.

What’s the context window and knowledge cutoff?

A 1,000,000-token context window with up to 128,000 output tokens, and a January 2026 knowledge cutoff. There’s no long-context price premium.

What was new in Opus 4.7 versus Opus 4.6?

Big coding gains (SWE-bench Verified 80.8% → 87.6%; SWE-bench Pro 53.4% → 64.3%), the first automatic cybersecurity safeguards (Project Glasswing), 3.75-megapixel vision, a new xhigh effort level, better file-system memory, task budgets (beta), and the /ultrareview Claude Code command.

Why are some of my prompts behaving differently on Opus 4.7?

Opus 4.7 follows instructions more literally than earlier models, so prompts that relied on looser interpretation can produce unexpected results. Anthropic recommends re-tuning prompts and harnesses when migrating.

Is Opus 4.7 the most powerful Claude?

No. At its launch, Claude Mythos Preview was more capable (but restricted). Since 28 May 2026, Opus 4.8 is the flagship, with the Mythos-class Fable 5 above it.

Can I use Opus 4.7 for cybersecurity work?

Yes, for legitimate purposes — but safeguards automatically block prohibited or high-risk requests. Security professionals doing vulnerability research, pen-testing or red-teaming can apply to Anthropic’s Cyber Verification Program.

Last verified 16 June 2026. Benchmark figures are Anthropic-reported via the Claude Opus 4.7 System Card unless otherwise noted; GPT-5.4 and Gemini 3.1 Pro figures are the best reported API versions per Anthropic’s charts. Pricing and availability current as of the publication date and subject to change.