Claude Sonnet 5.5 Jumped From 10.3% to 70.6% on Terminal-Bench — That Gap Is an Agent Upgrade Audit Business.

Anthropic didn't just ship a faster, cheaper model with Sonnet 5.5. It published proof that whatever agent workflows are running on Sonnet 5 today are leaving a measurable amount of capability on the table.

What Anthropic actually shipped

Released September 28, 2026, Sonnet 5.5 is pitched on speed and cost first: "30%+ faster than Sonnet 5" and "costs up to 30% less for most work." Pricing lands at "$2" per million input tokens, "$10" per million output tokens, "$0.20" for cache reads, and "$2.50" for cache writes.

But the benchmark deltas are where the audit business lives:

• Terminal-Bench 4.0: "70.6%" versus Sonnet 5's "10.3%"
• CursorBench 4.0: "55.5%" versus "34.1%"
• FrontierCode 1.1 (Main) at Max: "46.2%" versus "42.4%"
• OSWorld 2.1 (partial): "80.1%" versus "57.0%"
• Humanity's Last Exam (with tools): "64.5%" versus "54.9%"
• Chartography (no tools): "61.6%" versus "15.6%"
• GDPval-AA v2.1: "1844" versus "1449"
• AA-Briefcase v1.1: "1811" versus "1359"

It's also "the first Sonnet model to beat Pokémon Red using only screenshots" — a small detail, but it signals the computer-use and visual-agent gap closed hard in this release too.

Why this is a service, not just an upgrade

Sonnet 5.5 ships as a drop-in: the model ID is claude-sonnet-5-5, available now on "Claude Platform, AWS, Google Cloud, and Microsoft Azure." No new SDK, no new orchestration layer — teams already calling Sonnet 5 can point the same integration at the new model ID. That's what makes this auditable and sellable in the same week: there's no rebuild to sequence around, just a benchmark to run and a config value to change.

The moneyPlay in practice

1. Pitch a fixed-scope "Sonnet 5.5 upgrade audit": re-run a client's existing agent tasks against the new model and report the delta on the exact benchmarks Anthropic published, not a generic "it's better now" summary

2. Quantify the cost side first — the "30%+ faster" and "costs up to 30% less for most work" claims translate directly into a monthly token-spend number the client can verify against their own usage

3. Scope the migration as a config change: same API shape across all four platforms via claude-sonnet-5-5, so the engagement timeline is measured in days, not sprints

4. For clients running long-horizon, research, or computer-use agents, lead with OSWorld and Humanity's Last Exam deltas specifically — those are the workloads where the capability jump is largest and easiest to demonstrate live

5. Price the audit as a flat fee plus a share of the measured savings, since the entire pitch reduces to a dollar number pulled from Anthropic's own published pricing and benchmarks

Bottom line

Anthropic quantified its own model's agentic-coding weaknesses inside the announcement of the fix. The upgrade audit — measuring exactly how much of that Terminal-Bench, CursorBench, and OSWorld gap a client is still eating on Sonnet 5 — is a one-to-two week engagement nobody has to build from scratch to sell.

Sources:

https://www.anthropic.com/claude-sonnet-5-5