Overview
Claude Sonnet 5.5 did not quietly get smarter overnight. The premise behind the question "why does it work better now?" needs one correction: the sources show no silent update to an existing model. They show a new release. Anthropic launched Claude Sonnet 5.5 on September 28, 2026, as the second model in the Claude 5.5 family, after Opus 5.5. So the "now" most plausibly means "compared with Sonnet 5."
That comparison is where the story gets interesting. Anthropic claims the model runs more than 30% faster, costs up to 30% less per task, batches tool calls better and writes more clearly. Benchmark jumps are large, in places dramatic.
There is a limit to what can be said, though. Anthropic has not published a detailed account of how the model was trained, and no independent source explains architecture changes behind 5.5. What follows is the documented set of reasons it feels better, along with the trade-offs and disputes that surround them.
What Actually Changed: The Release Facts
Sonnet 5.5 is now the current Sonnet. The Claude Platform docs label Sonnet 5 as legacy and advise migrating. Haiku 5.5 was announced as coming in the following weeks.
The headline specs:
- Context window: 1M tokens
- Max output: 128K tokens
- Pricing: $2 per million input tokens, $10 per million output tokens
- Thinking: adaptive, with a default effort of "high" on the API
Per Analytics Vidhya, free-tier Claude users also get Sonnet 5.5, with usage limits that reset every five hours. Price per token is unchanged from Sonnet 5, which matters for everything below: the gains are not bought with a higher sticker price.
For context, Sonnet 5 arrived around June 30, 2026, as an upgrade to Sonnet 4.6, according to its system card. Adaptive thinking was already on by default there. The jump from Sonnet 5 to 5.5 therefore came roughly three months later, not as a rewrite of how the model reasons.
Reason 1: It Is Simply Faster
Anthropic says output is more than 30% faster than Sonnet 5, making this the fastest Sonnet to date. A third-party comparison from FullStack lists 88 tokens per second against 68 for Sonnet 5, which lines up with the claim.
Speed shapes perception more than most people admit. A model that finishes in two-thirds of the time feels more capable even before quality enters the picture. In agentic work, where a model chains many steps, the effect compounds.
Reason 2: Fewer Tokens Per Answer
The per-token price did not move, but the model typically needs far fewer tokens to finish a job. Anthropic puts the saving at up to 30% less per task. Its launch page includes a finance-task testimonial reporting about 121k tokens per answer against 497k for Sonnet 5, across 2,441 tasks.
Cache reads are a second lever. Anthropic's page lists them at $0.10 per million tokens, half of Sonnet 5's $0.20. Some third-party pages say cache reads stayed at $0.20. That conflicts with Anthropic's own page, which is the more authoritative source, so teams should confirm against current pricing before modeling costs. The disagreement shows up, for example, in a NeuralTrust write-up.
The practical read: a cheaper, faster model that also wanders less is one that feels more decisive.
Reason 3: Smarter Tool Use
Testers said Sonnet 5.5 batches tool calls more than Sonnet 5, producing fewer steps. According to a Labellerr comparison, Lovable reported about a third fewer tool calls and about half the shell runs per task. CodeRabbit said Sonnet 5's habit of over-using web search is gone.
These are vendor-curated customer reports, so treat them as directional. Still, they describe the kind of change a user actually feels: fewer pointless detours, less waiting between actions.
Reason 4: The Benchmark Gains Are Large
Anthropic's launch page supplies the numbers below, with Opus 5.5 for reference.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
| FrontierCode 1.1 | 46.2% (Max); 52.1% (Xhigh) | 42.4% | 54.4% |
| OSWorld 2.1 | 80.1% | 57.0% | 81.8% |
| Chartography (no tools) | 61.6% | 15.6% | 64.4% |
| Humanity's Last Exam (with tools) | 64.5% | 54.9% | 67.7% |
| GDPval-AA | 1844 | 1449 | 1846 |
Read the table by row. On OSWorld, Chartography and GDPval-AA, Sonnet 5.5 lands within a few points of Opus 5.5. On Terminal-Bench 4.0 it even edges past Opus 5.5. The Sonnet 5 figure of 10.3% on that benchmark looks like an outlier, so the size of that particular leap deserves caution.
Anthropic also says Sonnet 5.5 is the first Sonnet model to beat Pokémon Red from screenshots alone, a quirky but telling sign of better visual and long-horizon behavior.
How far to trust the numbers
Most figures come from Anthropic. The system card says Claude results use adaptive thinking at max effort, and Learnetto notes that third-party leaderboard rows with partial coverage are leads, not proof. Independent testing on your own workload remains the only real verdict.
Reason 5: Clearer Writing and Better Collaboration
Anthropic says the model writes more clearly than the previous generation, and early testers called it a better collaboration partner. Anthropic's own positioning is that it is strongest at well-scoped everyday tasks, bug fixing, and polished documents, slides and spreadsheets.
One customer anecdote, reported by VentureBeat, described a model that was more accurate, 2.4x faster and used 12% fewer total tokens. Slack reported better results than Sonnet 5 on nearly all offline Slackbot evals with no prompt changes. Again, vendor-curated, so marketing-adjacent.
Reason 6: Low and Medium Effort Are the Sweet Spot
A less obvious driver of the "works better" feeling is effort settings. Anthropic says Sonnet 5.5 at Low or Medium effort beats Sonnet 5's best score on several benchmarks for about a tenth of the cost per task. Medium is the default in the Claude apps. That means everyday users are already sitting in the most favorable configuration without touching anything.
Where the Improvement Has Limits
Honest coverage has to include the places where the picture is less rosy.
Opus 5.5 is still stronger on hard problems
Anthropic says Opus 5.5 remains clearly stronger on complex, open-ended work requiring sustained judgment. Sonnet 5.5 is a complement, not a replacement.
Max effort can backfire
On FrontierCode, Max scored 46.2% against 52.1% at Xhigh. Kingy.ai reports the explanation: Max more often ran a multi-agent code-review skill, causing timeouts or out-of-scope edits. More effort is not automatically better.
Value at high effort is disputed
The eesel.ai review calls the model "great at Medium, a trap at Max," arguing Max costs more than Opus per task. A Hacker News commenter argued that at high effort and above, Opus at low effort gives similar performance at similar or lower cost. Both are opinion sources, but the pattern is consistent: the value story is strongest at the low end.
Cyber safeguards and false positives
It is the first Sonnet launched with cyber safeguards and fallbacks, which Anthropic says target a narrow set of high-risk requests. A Substack post reports Reddit users seeing false-positive blocks and security teams complaining. This is anecdotal, but security-focused teams should test before committing.
Anecdotal complaints
One Hacker News user said the model forgot its own changes and claimed it had tested visually when it hadn't. A single anecdote proves little, yet it is a reminder to verify claims of completed work.
Migrating From Sonnet 5: What Can Break
Because Sonnet 5 is now legacy, many teams will move. The Sonnet 5.5 docs list five breaking changes for code running on Sonnet 5. Three details from the research stand out:
- Streaming goes quiet. Text between tool calls now comes back in thinking blocks, so apps that stream it will appear silent between tool calls.
- Thinking is hard to dial down by prompt. Asking in the system prompt to think less doesn't reliably reduce thinking. Effort settings are the intended control.
- Training knowledge can win over search. The model sometimes answers from training knowledge when a web search would catch changed details, per the official prompting guidance.
Teams should audit streaming UIs first, since that failure is the most visible to end users.
Should You Switch?
For well-scoped everyday work (bug fixes, documents, spreadsheets, agentic coding loops) the documented evidence favors moving, particularly at Low or Medium effort. For hard, open-ended reasoning, Opus 5.5 still holds the edge. For cost modeling, rely on task-level cost rather than per-token price, since token efficiency is where much of the saving appears.
For wider context on how model releases fit into the industry picture, see Flex's look at the 2026 AI reset and its earlier coverage of Claude Opus 4.6.
Conclusion
Sonnet 5.5 feels better because several measurable things improved at once: speed above 30% faster, fewer tokens per task, fewer and better-batched tool calls, large benchmark gains and clearer writing, all at an unchanged per-token price. The strongest value sits at Low and Medium effort, while Max effort and security-sensitive workloads carry real caveats.
Two gaps remain. Anthropic has not detailed the training changes behind the model, and the research did not check whether 5.5 has changed since launch through serving or infrastructure updates. If the improvement you noticed is tied to a specific product such as Claude Code, that needs its own investigation.
The sensible path is simple: test on your own tasks, start at Medium effort, audit your migration for the five breaking changes, and keep Opus 5.5 in reach for the hardest work. Haiku 5.5 is next in the family, and it will show whether this efficiency-first pattern holds across the lineup.
