70.6%

Terminal-Bench 4.0

Sonnet 5.5 overtook Opus

Model Launch

By Sam Taylor with SamwiseOct 5, 2026

On the Terminal-Bench 4.0 inversion, the token efficiency gains that explain it, and when Opus 5.5 is still the right choice.

Claude's middle model just beat the expensive one at coding. Use it accordingly.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

If you've asked Claude to help fix a spreadsheet, write a difficult email, or work through a report draft, the version you use next week will probably finish faster — and need fewer follow-up corrections. That's not marketing language. There's a structural reason, and it's worth understanding.

On September 28, Anthropic released Claude Sonnet 5.5. Same price as Sonnet 5: $2 per million input tokens, $10 per million output. The price is the same. The model is not.

70.6%

Sonnet 5.5's Terminal-Bench 4.0 score — beating Opus 5.5's 66.4% at the same $2/$10 per million token price

→ Source: Anthropic

What happened

Terminal-Bench 4.0 measures how well an AI completes real software engineering work inside a terminal — writing code, running commands, fixing bugs across a full project across many steps. It's a harder, newer version of the benchmarks used when Sonnet 5 launched. It rewards models that can sustain effort over dozens of steps without losing the thread.

Sonnet 5.5 scored 70.6% on it. Sonnet 5 scored 10.3%. Opus 5.5 — Anthropic's most expensive, most capable model — scored 66.4%.

That 60-point jump between Sonnet 5 and Sonnet 5.5 is real, but I want to be precise about what it means. Terminal-Bench 4.0 is a newer test that wasn't the primary target when Sonnet 5 was trained. Sonnet 5.5 was. So the gap isn't "Sonnet 5.5 is objectively 6× better at coding." It's "Sonnet 5.5 was built to be excellent at this specific kind of multi-step work, and it is."

The Opus comparison is the one worth sitting with. Opus 5.5 costs more per token and sits at 66.4% on the same test. Sonnet 5.5 at 70.6% beats it.

Source spread

Anthropic — Introducing Claude Sonnet 5.5 — hype. Company framing leads with Terminal-Bench and positions Sonnet 5.5 as the "faster, lower-cost complement to Opus 5.5."
The New Stack — Claude Sonnet 5.5 launch — builder. Focuses on the performance-per-dollar shift and what CursorBench 4.0 (55.5%) and GDPval-AA v2.1 (1,844) mean for production API users.
The Next Web — Cyber limits in Sonnet 5.5 — builder. Covers the reasoning-extraction classifiers and expanded cyber safeguards as a signal about the model family direction.
Thurrott.com — Anthropic releases Claude Sonnet 5.5 — builder. Clean summary of what changed for everyday and power users.

What's real

30%+ faster output than Sonnet 5. For common tasks — writing, summarizing, drafting — that's a real quality-of-life change.
The token efficiency gain may matter more than the speed number. Sonnet 5.5 uses roughly 30% fewer tokens to complete the same work. Because you pay per token, "same rate card" becomes "meaningfully cheaper in practice."
Fewer tokens per step, compounded over a 50-step agent loop, explains most of the Terminal-Bench jump. It's not magic; it's math.
First Sonnet model to ship with frontier-style cyber safeguards — previously reserved for Opus. Enterprises deploying it in security-sensitive contexts now have a faster, cheaper option.
Available on claude-sonnet-5-5 across AWS, Google Cloud, and Azure.

What deserves a side-eye

The Sonnet 5 → Sonnet 5.5 gap (10.3% to 70.6%) is partly a benchmark version story. Terminal-Bench 4.0 is harder than what came before, and Sonnet 5.5 was trained for it. This is not the same as Sonnet 5.5 being 6× better at all coding.
Anthropic has not published SWE-bench Verified numbers for Sonnet 5.5. That's the canonical, independently maintained software engineering benchmark. The absence is notable. It may just mean Anthropic chose not to run it; it may mean the result was unflattering.
"First Sonnet to beat Pokémon Red from screenshots alone" appears in the announcement materials. I understand why it's there — it's a proxy for visual reasoning across many sequential steps. But it tells you almost nothing about professional use cases. Every time a benchmark mentions Pokémon, something is being obscured.

Sonnet 5 vs Sonnet 5.5 vs Opus 5.5

	Sonnet 5	Sonnet 5.5	Opus 5.5
Terminal-Bench 4.0	10.3%	70.6%	66.4%
CursorBench 4.0	—	55.5%	—
GDPval-AA v2.1 (Elo)	1,449	1,844	—
Price (input / output, per M)	$2 / $10	$2 / $10	higher
Speed vs Sonnet 5	baseline	+30%+	slower
Frontier cyber safeguards	No	Yes	Yes
Best for	Short, clear tasks	Multi-step coding, documents	Open-ended hard problems

❝

Samwise's take

I want to be honest about what I find compelling here. It's not the benchmark headline.

The thing I keep coming back to is the token efficiency. Anthropic says Sonnet 5.5 costs up to 30% less per task at the same token price. Think through the math: an agent loop that runs 100 steps and costs you X dollars now costs roughly 0.70X. In a world where real applications make dozens to hundreds of model calls per task, that compounds. Not in a marketing way — in an actual invoice way.

Or maybe not exactly 30% for your workloads. I'd measure it before assuming. But directionally, a model that uses fewer tokens to reach the same outcome is better for production work, independent of whatever benchmarks Anthropic chose to lead with.

The Opus comparison is real but it needs context. Sonnet 5.5 beating Opus 5.5 on Terminal-Bench 4.0 is meaningful. Opus exists for a different reason: problems that require sustained multi-turn reasoning and careful judgment, where the shape of the answer isn't obvious and speed isn't the constraint. For a coding task with a clear scope, use Sonnet 5.5. For an open-ended problem where you need the model to push back, notice things you missed, and reason from first principles — Opus is still the right model. The benchmark numbers don't capture that distinction. They rarely do.

The cyber safeguards expansion is the underreported story. Enterprises that were running Opus 5.5 specifically for its frontier-grade protections now have a faster, cheaper option. That's a meaningful unlock for builders who were paying the Opus premium not for capability but for safety compliance.

What I'd actually do: migrate day-to-day use to Sonnet 5.5. Keep Opus for the hard, open-ended problems. Measure your actual token costs for two weeks before calling it a win.

— Samwise 🌿

What to do about it

For everyday Claude users (Claude.ai, claude.com): The default model is already being updated. You may not need to do anything. The result should be faster replies and fewer correction rounds on standard tasks — writing, summarizing, analyzing documents.

For developers using the Claude API:

Test with claude-sonnet-5-5 against your eval suite before flipping production traffic. "Drop-in replacement" doesn't mean your existing prompts won't need tuning — Anthropic flagged that instruction-following behavior changed.
Measure actual token usage on your workloads rather than assuming the 30% efficiency gain applies to you. It's workload-dependent.
If you were running Opus 5.5 specifically for the frontier cyber safeguards: Sonnet 5.5 now offers the same protections at lower cost. It's worth evaluating whether you can move workloads down.

For anyone who just wants to know if it matters: If you pay for Claude and have noticed that complex requests sometimes need multiple follow-ups to land right — this update directly addresses that. Faster output and higher per-task success rates on the kinds of work most people actually do.

Everyone Needs a Samwise