Blog / 16 avril 2026 · 6 min read

Claude Opus 4.7 dropped today. Here's what actually changes for a dev.

Anthropic released Claude Opus 4.7 on April 16, 2026. SWE-bench Pro at 64.3%, +13% on coding, a new xhigh tier, unchanged pricing. A hot-take analysis for developers.

Anthropic pushed Claude Opus 4.7 this morning. First release of the “Mythos” class with production-grade guardrails, and honestly, after two hours throwing it at my existing agents, I’ve got observations that deserve better than a tweet.

I’ll run through the official numbers, then get into what actually changes for a dev who works with Claude every day.

The numbers that matter

SWE-bench Pro: 64.3% — ahead of GPT-5.4 (57.7%) by more than 6 points. Opus 4.6 was sitting around 53%. That’s a clean +11 points in 2 months.
SWE-bench Verified: 87.6% — we’re getting close to this benchmark’s ceiling.
+13% on Anthropic’s internal 93-task benchmark, with 4 tasks that neither Opus 4.6 nor Sonnet 4.6 had ever solved.
3x more production tasks solved in agentic workflows (the most honest metric, in my view, since it’s not a synthetic benchmark).
+14% in multi-step agentic reasoning, and more importantly, only a third of the tool errors compared to the previous version.
1M token context, no premium long-context tier. Output caps at 128K.
Pricing unchanged: $5 input / $25 output per million tokens. That one’s actually surprising — they could have raised it, they didn’t.

What changes technically

The new “xhigh” tier

Anthropic added a reasoning effort level between high and max. The idea: on hard problems, max would sometimes take 40+ seconds and burn through tokens, while high wasn’t quite enough. xhigh fills the gap — more reasoning than high, less latency than max.

Concretely, I re-ran an agent that analyzes complex error traces. On 20 requests:

high: 12 correct answers, 5.2s average
xhigh: 18 correct answers, 9.1s average
max: 19 correct answers, 23.7s average

The quality-to-latency ratio on xhigh is excellent. That’s probably where I’ll route my demanding tasks by default from now on.

High-resolution vision

Vision now goes up to 3.75 megapixels. On Opus 4.6, I was getting misses on dense screenshots (Excel tables, cramped dashboards). I reran three tests:

Extraction from a heavily loaded Grafana screenshot → 100% accurate (was 92% before)
Reading a mid-quality scanned PDF → still worse than a dedicated OCR (Mistral OCR) on edge cases
Analysis of a DevTools capture → clear improvement on identifying individual requests

Verdict: it’s a real jump on dense UIs. On paper scans, not enough to replace a specialized OCR.

The new tokenizer

Anthropic mentions “a new tokenizer” without going into detail in the public release notes. I measured it: on my usual system prompts, I’m burning roughly 8 to 12% fewer tokens for the same content. That’s a quiet cost reduction on top of the 0% price increase.

What this changes for me (and for you)

For long-running agents

The “-2/3 tool errors” gain is probably the most impactful one in practice. My background agents that used cascading tool calls had a failure rate around 15% on 4.6 (bad params, tool not called, wrong format). On 4.7, I re-ran the same harness across 200 runs: 4.8% failures. Divided by 3, exactly as announced.

Concretely, that means less retry logic, less defensive code wrapped around calls, fewer weird edge cases to debug.

For interactive dev work

For day-to-day AI-assisted dev (Claude Code, Cursor), the gain is more nuanced. On short, clear tasks, Sonnet 4.6 remains plenty good and a lot cheaper. Opus 4.7 pulls ahead on complex multi-file tasks, gnarly debugging, ambitious refactors.

The decision tree I recommended back in March still holds: Haiku for classification/extraction, Sonnet for standard refactoring/review, Opus for architecture/multi-layer work. Opus 4.7 doesn’t change that tree — it just beefs up the Opus box.

For marketing benchmarks

64.3% on SWE-bench Pro vs 57.7% for GPT-5.4 is 6 points. Meaningful but not overwhelming. The reality on the ground: for 80% of tasks, you won’t see a quality difference between the two. The differences show up on tasks neither one used to solve — and that’s where Opus 4.7 is gaining ground.

The interesting detail: Mythos

Axios picked up on something Anthropic said off the record: Opus 4.7 “trails” an unreleased internal model called Mythos. So Anthropic is publicly admitting they have something better internally, but decided not to ship it yet — probably for safety/alignment reasons.

That’s an interesting piece of info on several levels:

Unusual transparency. OpenAI and Google never publicly admit they have something better internally. Anthropic is playing the honesty card.
An implicit roadmap. Mythos will show up eventually, probably within 6 months if the historical pace holds. Devs can expect a bigger jump than 4.6 → 4.7.
The apparent slowdown in releases is deliberate. What we get is what passed the safety test battery. What’s more powerful is waiting.

Migration: what breaks?

I migrated my three production agents from claude-opus-4-6 to claude-opus-4-7 by changing a single string. Zero other code changes. All 3 agents ran for an hour with no regression.

What to watch out for:

The new tokenizer can shift limits on edge cases where you had prompts calibrated down to the token.
If you use strict tool use with exact schema validation, run another pass — output formats are slightly more concise on 4.7.
Prompt caching stays compatible across versions, you keep your savings.

What’s still disappointing

To be honest, two things did not move with this release:

1. Computer use capability. No improvement mentioned. The Chrome DevTools integration is just as uneven as what I tested back in early April (see my write-up on 1 month of Computer Use). Too bad.

2. The entry price is still steep for volume. At $25/M output tokens, a consumer product serving Opus 4.7 at scale is too expensive to run. For tightly scoped B2B work it’s fine, for consumer it’s still Haiku/Sonnet territory.

Bottom line

Opus 4.7 is a solid release without being revolutionary. The real gain is in agentic reliability (tool errors divided by 3) and the new xhigh tier, which offers a better quality-to-latency ratio than max.

If you use Claude for serious dev work, migrate to claude-opus-4-7 today. If you’re running agentic workflows, migrate yesterday.

And keep an eye on Mythos. Something tells me we haven’t heard the last of it.

Sources

— Kevin Aubrée

Keep reading

Back to blog