Claude Opus 4.7 vs Opus 4.6

Anthropic released Claude Opus 4.7 on April 16, 2026, two months after Opus 4.6. It's a same-tier upgrade — not a new model family — and the pitch is specific: same price, higher capability per token, longer autonomous runs. This post is the direct head-to-head. Every benchmark, every price, every behavior change, and a migration checklist you can use as a reference when flipping the model flag in production.

The Verdict

Opus 4.7 is a drop-in upgrade to Opus 4.6. It beats 4.6 on 12 of 14 reported benchmarks, adds a new xhigh effort level, sees images at 3.3× higher resolution, follows instructions more literally, and introduces self-verification on long-running agentic work. The per-token price is unchanged. Low-effort 4.7 matches medium-effort 4.6 on quality, so the effective cost per completed task drops even though the $5 / $25 rate is identical.

Two caveats worth knowing before migrating. First, the updated tokenizer can map the same text to 1.0–1.35× more tokens than 4.6, which affects budgeting. Second, because 4.7 takes prompts more literally, prompts tuned for 4.6's looser interpretation can misfire. Both are quick to audit but worth a pass.

Side-by-Side at a Glance

Nothing on the commercial surface changed: same context window, same pricing tiers, same platforms. The differences are operational — new effort level, updated tokenizer, higher-resolution vision, and new behaviors around self-verification and instruction following.

Benchmark Deltas

Every benchmark below is self-reported by Anthropic in the Opus 4.7 launch announcement. The chart sorts by delta so the largest moves appear first. Positive bars (in Anthropic orange) are gains; negative bars (in rose) are regressions.

The story on the benchmark page is that gains concentrate on the hardest, least-saturated problems: SWE-bench Pro (+10.9pp) jumps more than SWE-bench Verified (+6.8pp), HLE without tools (+6.9pp) jumps more than HLE with tools (+1.6pp), and MCP-Atlas (+14.6pp) — the agentic tool-use benchmark — takes the single largest jump of the release.

Coding

Partner numbers reinforce the benchmark story. Replit reports same-quality output at lower cost, Rakuten measured 3× more production tasks resolved, and Cursor reports 70% on CursorBench vs 58% for Opus 4.6.

Reasoning & Knowledge

Agents

BrowseComp is the only real regression. Opus 4.6's 84.0% was measured under a multi-agent harness at max effort; the comparison is sensitive to harness choice, but the delta is honest enough to flag. CyberGym is effectively flat by design — Anthropic states it "experimented with efforts to differentially reduce" offensive cyber capabilities during training.

Vision

The no-tools delta is the one that matters: +13.4pp isolates the vision capability itself rather than the image-cropping tool. XBOW similarly reports a 98.5% visual-acuity score on Opus 4.7 (vs 54.5% on 4.6) — large enough to unblock autonomous pen-testing workflows that weren't viable on 4.6.

Win / Loss Scorecard

Grouped by capability domain. Cybersecurity is called out separately because it's a deliberate non-improvement rather than a regression.

What Actually Changed

Benchmarks are the score; behaviors are the story. Four operational changes make Opus 4.7 feel different in practice, even on prompts where the benchmark delta is small.

Self-verification before reporting

Anthropic's framing: Opus 4.7 "devises ways to verify its own outputs before reporting back." In practice this means the model writes tests, runs sanity checks, and inspects its own output before declaring a task complete. Vercel reports 4.7 "does proofs on systems code before starting work" — behavior not seen on 4.6. On long agentic runs, this is the single change most users report feeling first: fewer confident-but-wrong reports back.

Literal instruction following

Opus 4.7 follows instructions more literally than 4.6 or any prior Claude model. Anthropic flags this explicitly as a migration concern: prompts that depended on loose interpretation may now produce unexpected results because 4.7 takes the wording at face value. The most common failure mode is bullet lists of "suggestions" that 4.6 treated as optional hints being read as hard requirements on 4.7. Audit system prompts before rollout.

Higher-resolution vision

Images up to 2,576 pixels on the long edge (~3.75 MP) — versus ~1,568 px (~1.15 MP) on 4.6. 3.3× more pixel area per image, applied automatically through the vision API. Two practical consequences: computer-use agents can read dense screenshots without operator-side pre-cropping, and data extraction from complex diagrams jumps sharply (see the +13.4pp CharXiv-R no-tools delta).

File-system memory for multi-session work

Opus 4.7 is better at reading, writing, and reusing notes on a persistent file system across sessions. For agents that work over days rather than minutes — think a long-running engineering task that spans multiple model turns across multiple sessions — this removes the need to re-establish context at the start of every run.

New xhigh effort level

xhigh is a new tier between high and max. It gives developers finer control over the reasoning-vs-latency tradeoff: more thinking than high, without the full cost of max. Claude Code raised its default effort to xhigh for all plans on the 4.7 rollout.

Cost: Same Price, Fewer Tokens

Per-token pricing is identical. The cost story lives at the task level: Hex's early-access testing found low-effort Opus 4.7 matches medium-effort Opus 4.6 on quality, so the same completed work uses meaningfully fewer tokens at a cheaper effort tier. Anthropic's internal coding eval similarly reports token usage per completed task improved at every effort level — accuracy rises faster than token spend.

There are two offsetting pressures to plan for. The updated tokenizer can map the same text to 1.0–1.35× more tokens than 4.6, so static token budgets built on 4.6 need re-measuring. And at higher effort levels, 4.7 thinks more than 4.6 did — output token counts can rise on reasoning-heavy turns even when the task itself is the same. On aggregate, these effects are smaller than the quality-per-effort-tier gain, but any individual workload should be benchmarked before flipping the flag at scale.

Migration: 4.6 → 4.7

Opus 4.7 is API-compatible with 4.6. The surface change is the model ID (claude-opus-4-6 → claude-opus-4-7). Everything else is an evaluation exercise, not a code change.

Checklist

• Audit system prompts for literal-instruction risk. Hunt for phrases like "consider," "you might," and bullet lists of "suggestions." 4.7 reads these closer to hard requirements than 4.6 did.
• Re-measure token budgets on real traffic. The tokenizer change can shift the same text by up to 35%. Static cost forecasts built on 4.6 will drift; dynamic budgets that meter on-the-fly are unaffected.
• Pick an effort level intentionally. For coding and agentic use cases, start with high or xhigh. For simple classification or summarization, low on 4.7 typically matches medium on 4.6 at lower cost.
• Turn on task budgets if you run agents. Paired with xhigh, budgets let you say "think hard on this, but don't burn more than N tokens finishing it." Prevents runaway multi-agent branches.
• Downsample images only when needed. Opus 4.7 processes higher-res images automatically, which costs more tokens per image. Workloads that don't need the extra detail should downsample client-side before sending.
• Run an eval pass. Especially for BrowseComp-style browsing under a specific multi-agent harness, where 4.7 can regress. A quick A/B on a held-out set catches these before production.
• Apply to the Cyber Verification Program if relevant. Legitimate offensive-security work (vulnerability research, pen-testing, red-teaming) on 4.7 requires the new program to avoid default refusals. Not needed for defensive or general-purpose workloads.

Anthropic's official migration guide includes concrete prompt-tuning advice and token-usage comparisons per effort level.

When to Upgrade, When to Stay

For the full announcement, see Anthropic's Opus 4.7 launch post and system card. For a deeper breakdown of the release itself, see our Opus 4.7 launch overview.