AI ENGINEERING / SEPTEMBER 2026 · 7 MIN READ
Claude Sonnet 5.5: Faster, Cheaper—What to Move Now
Anthropic has introduced Claude Sonnet 5.5, a faster, lower-cost complement to Opus 5.5. It’s 30%+ faster and can cost up to 30% less per task than Sonnet 5, while posting strong results on agentic coding and long-horizon work, according to Anthropic. Here’s how I’d decide what to migrate now—and how to measure it.
IN THIS ARTICLE What just happened · Why it matters for teams · Migration candidates: where Sonnet 5.5 fits · A practical LLM benchmarking playbook · Migration playbook for Claude Sonnet 5.5: what to do this week
KEY TAKEAWAYS
01 Move well-scoped coding, bug-fixing and doc/spreadsheet generation to Sonnet 5.5 first.
02 Benchmark at multiple effort levels and choose for capability-per-dollar, not raw score.
03 Keep complex, open-ended judgement work on Opus 5.5 for now.
What just happened
Claude Sonnet 5.5 is out, the second model in the Claude 5.5 family. It runs 30%+ faster than Sonnet 5 and, in Anthropic’s own testing, costs up to 30% less per task because it typically needs fewer tokens to do the same work, while prices remain $2 per million input tokens, $10 per million output tokens and $0.20 per million cache reads, per Anthropic. The model is positioned as a faster, lower-cost complement to Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides and spreadsheets, with a sharp eye for design, according to Anthropic.
On performance, Sonnet 5.5 posts 70.6% on Terminal-Bench 4.0—an agentic coding evaluation—compared to Sonnet 5’s 10.3%, is two points below Opus 5.5 on GDPval-AA, and shows strength on long-horizon work and image understanding (it’s the first Sonnet to beat Pokémon Red from screenshots), per Anthropic. Anthropic’s cost-versus-accuracy plots also show that at Low/Medium effort, Sonnet 5.5 can beat Sonnet 5’s best scores at about a tenth of the cost per task, and at higher effort settings it can perform comparably at similar cost, per Anthropic.
On alignment and safety, Sonnet 5.5 improves on or matches Sonnet 5 on most measures in Anthropic’s automated audit. With cybersecurity capabilities comparable to Opus 5, it’s the first Sonnet to launch with cyber safeguards and fallbacks like those used for Anthropic’s most capable models, while biology safeguards remain the same as Sonnet 5; safeguards target a narrow set of high‑risk requests and routine software development is unaffected, according to Anthropic.
Why it matters for teams
Speed and cost change the calculus for day‑to‑day engineering. If a model ships 30%+ faster and can reduce task costs by up to 30% relative to its predecessor while maintaining strong capability on relevant benchmarks, you get to iterate more often and broaden coverage without increasing spend—both stated by Anthropic. That’s exactly where Claude Sonnet 5.5 is pitched: well‑scoped tasks, bug‑fixing loops, and polished artefacts like docs, slides and spreadsheets.
There’s also a clear division of labour. Anthropic’s own testing and external testers found Opus 5.5 remains stronger for complex, open‑ended work requiring sustained judgement, while Sonnet 5.5 complements Opus 5.5 best at lower effort settings due to its cost per task profile, per Anthropic. In other words, treat Sonnet 5.5 as a throughput engine, not a universal replacement.
Finally, the safeguards story matters. Being the first Sonnet to ship with cyber safeguards and fallbacks like Opus 5’s—with routine software development unaffected—lowers friction for enterprise adoption without slowing everyday engineering flows, according to Anthropic. That’s a useful data point for enterprise AI risk reviews.
AYAN’S TAKE
I’d route well-scoped work to Sonnet 5.5 immediately and reserve the heavy judgement calls for Opus 5.5.
Migration candidates: where Sonnet 5.5 fits
Agentic terminal coding and code-change proposals are prime candidates. Sonnet 5.5’s 70.6% on Terminal‑Bench 4.0 and its improved showing on FrontierCode v1.1 indicate a material jump over Sonnet 5 for real‑world coding workflows, per Anthropic. If your current flow leans on the model to diagnose and patch within a constrained repo or CLI context, Sonnet 5.5’s speed and capability should translate into higher throughput and lower cost per merged change.
The deployment decision behind this is covered in Private LLM for Enterprise: When It Is Worth It.
Ambiguous, multi‑file tasks also look promising. On CursorBench 4.0, Sonnet 5.5 improves substantially over Sonnet 5, and Anthropic’s cost‑versus‑accuracy charts show stronger capability per dollar at lower effort levels, per Anthropic. Faster iteration helps here: you can afford more attempts without blowing the budget.
Long‑horizon knowledge work with clearly defined deliverables is another fit. On AA‑Briefcase and GDPval‑AA, Sonnet 5.5 lands close to Opus 5.5 in capability while costing less per task at lower effort settings, according to Anthropic. That’s a good signal for research summaries, structured briefs or stepwise planning—so long as the end state is crisp.
If you rely on visual understanding, Sonnet 5.5’s improvement on chart recognition (Chartography) without tools and its stronger image‑grounded behaviour (including the Pokémon Red result) are encouraging, per Anthropic. I’d pilot tasks like chart interpretation or layout‑aware document QA where errors are easy to spot and correct.
And don’t overlook collaborative drafting. Anthropic notes Sonnet 5.5 writes more clearly than the previous generation and is well suited to fast iteration on less complex tasks; it’s explicitly called out for creating polished documents, slides and spreadsheets, per Anthropic. That’s low‑risk, high‑leverage surface area for immediate migration.
What not to move first: complex, open‑ended work requiring sustained judgement. Anthropic is explicit that Opus 5.5 remains stronger there, and Sonnet 5.5 complements it best at lower effort where cost per task is advantageous, per Anthropic. If your primary constraint is raw cost at extreme scale, note that Claude Haiku 5.5, built for high‑volume and cost‑sensitive applications, will join the 5.5 family in the coming weeks, according to Anthropic—worth keeping in mind for model selection.
A practical LLM benchmarking playbook
Define success by workload, not by model. For each task class (e.g., repo‑scoped bug fixes, CLI agents, research briefs, chart interpretation), set an acceptance bar in three dimensions: quality (domain‑specific pass/fail and error taxonomy), speed (time to first token and time to final), and cost (tokens in/out and cost per task using the prices Anthropic lists). This mirrors Anthropic’s capability‑per‑dollar framing across effort levels, as described in their post.
If you are running systems like this in production, an AI architecture review tests retrieval, evaluation, permissions and cost before they become incidents.
Sample realistic prompts. Build a representative set that captures your data shapes: presence/absence of tools, image inputs for visual tasks, and ambiguity levels. Keep prompts constant across models so the only variables are the model and, intentionally, the effort setting. Anthropic’s results show Sonnet 5.5 shines at Low/Medium effort for cost per task, but can perform comparably at higher settings when needed—measure both for your workloads, per Anthropic.
Run controlled trials against your incumbent and Sonnet 5.5 at two or three effort levels. Record: latency distribution, tokens in/out, total cost per task, and outcome quality. For coding, borrow ideas from the benchmarks Anthropic cites (Terminal‑Bench, FrontierCode, CursorBench) by using merge‑ability or test pass rate as your quality proxy; for knowledge work (AA‑Briefcase/GDPval‑AA‑style), use rubric‑scored deliverables. The goal is to select the best capability per dollar for each workload.
Decide routing and fallbacks. Based on your thresholds, route well‑scoped, lower‑judgement tasks to Sonnet 5.5 by default. Escalate to Opus 5.5 when quality falls short or when the task is complex and open‑ended (aligning with Anthropic’s guidance that Opus 5.5 remains stronger there). If your domain overlaps with sensitive areas, note Sonnet 5.5’s cyber safeguards and fallbacks, with routine software development unaffected, per Anthropic—factor that into your evaluation notes rather than discovering it mid‑integration.
Track cost drift and tune effort. Anthropic’s charts show meaningful trade‑offs across effort levels; small increases in effort can raise cost per task but improve accuracy, while at Low/Medium effort Sonnet 5.5 can beat Sonnet 5’s best scores for roughly a tenth of the cost per task, according to Anthropic. Periodically re‑sample to ensure your choice remains optimal as prompts, datasets and team practices evolve.
Migration playbook for Claude Sonnet 5.5: what to do this week
Shortlist candidates. Move well‑scoped coding agents, repo‑bounded bug fixing, chart/document interpretation, and collaborative drafting of documents, slides and spreadsheets to a Sonnet 5.5 pilot first—these are use cases Anthropic highlights for the model in their announcement. Keep complex, open‑ended tasks on Opus 5.5 for now, per Anthropic.
Benchmark with intent. For each candidate workload, run A/B tests between your incumbent (whether Opus, GPT or current Sonnet) and Sonnet 5.5 across at least two effort settings. Judge on capability per dollar, not just headline accuracy. Sonnet 5.5 is 30%+ faster than Sonnet 5 and can cost up to 30% less per task in Anthropic’s testing, so expect iteration speed gains; validate the actual savings for your prompts using the published pricing in Anthropic’s post.
Roll out with guardrails. Start with routing rules that prefer Sonnet 5.5 for well‑scoped tasks and escalate to Opus 5.5 when your quality checks fail or the task is inherently open‑ended. Monitor latency, costs and error types for two weeks, then freeze the routing policy until the next evaluation cycle. If you’re heavily cost‑sensitive at very high volumes, plan a parallel track to evaluate Claude Haiku 5.5 when it arrives in the coming weeks, per Anthropic.
Bring stakeholders with you. Document the evaluation method and results as part of your LLM benchmarking notes; it will help internal reviewers, procurement and risk understand the trade‑offs. Sonnet 5.5’s alignment profile and cyber safeguards (with routine software development unaffected) are useful talking points from Anthropic for enterprise AI governance conversations.
My view: if a task is well‑scoped and measurable, move it to Sonnet 5.5 now and bank the iteration speed; if it’s open‑ended and judgement‑heavy, keep it on Opus 5.5 and revisit once you’ve got solid evidence that Sonnet’s higher‑effort settings meet your bar.
Read alongside How to Test an AI Vendor’s Claims Before You Sign or Invest and LLM deception testing: a pragmatic release-gate playbook.
READ NEXT
GENERATIVE AI & LLMS / THE PRACTICAL CTO
Private LLM for Enterprise: When It Is Worth It
Private LLM for enterprise, honestly assessed: the four deployment models, the guarantees each one actually buys, the total cost people underestimate, and a decision checklist.
GENERATIVE AI & LLMS / THE PRACTICAL CTO
How to Test an AI Vendor’s Claims Before You Sign or Invest
AI claims are easy to make and hard to verify. Five questions that separate real capability from a demo, whether you are buying the product or investing in the company.
