Claude Sonnet 5.5: More Thinking, More Cost, No Image, and Why the License Comes Before the Benchmark

The news items of September 30, 2026 raise three questions for anyone running AI agent setups, and all three end on the same lesson: The number that fires first is not the number that counts. Item one: Anthropic has released Claude Sonnet 5.5, and the interesting part is not the benchmark headline but the effort jump. Item two: The company behind the new Holo 4 release shipped two agent models that can operate a computer, and the real decision sits not in the benchmark score but in the license text. Item three: A study with email agents shows that the best tested configuration passes only 33.5 percent of scenarios even though 99.7 percent of the tool calls ran without errors. The advice of these items: First sort out the license, then nail down where a human signs off before money moves or a message goes out. Only then does comparing percentages pay off.

Simon Willison had the model draw a pelican on a bicycle as an image, comparing three effort levels. At the Xhigh level, a result arrived after 41 seconds for 5.74 cents. At the highest level Max, the model thought for 128,000 tokens, that cost $1.28, and still no image came out. The price for the unusable result sits at roughly twenty-two times the usable one. This is not a curiosity at the margin, it is the operating principle: With Sonnet 5.5, the thinking effort keeps growing linearly, but the hit rate does not grow with it.

And yet Sonnet 5.5 is supposed to be the budget model. The price stays identical to its predecessor: $2 per million input tokens, $10 per million output tokens, cache reads at $0.20. The saving effect does not live in the token price but in tokens per task: Anthropic states the model needs up to 30 percent fewer tokens per task than Sonnet 5 and outputs more than 30 percent faster. Anthropic's own numbers also show that more thinking does not automatically mean better: On the coding benchmark FrontierCode 1.1, Sonnet 5.5 reaches 52.1 percent at the Xhigh level but only 46.2 at Max. Anthropic's explanation: At Max, the model more often calls a code-review function that spreads the work across several subagents, and in examined cases this caused timeouts or edits outside the task's scope, both of which the benchmark penalizes. Anthropic itself recommends Opus 5.5 as the better choice for complex, open-ended tasks.

The rest of the map from the announcement: Terminal-Bench 4.0, the test for agentic coding on the command line, jumps from 10.3 percent for Sonnet 5 to 70.6 percent, which sits even above the best score of Opus 5.5, measured there at 66.4 percent at the Xhigh level. On CursorBench 4.0, the recreation of real Cursor coding sessions, it is 55.5 against 34.1 percent for the predecessor, only two points below Opus 5.5 at 57.8. In knowledge work, Sonnet 5.5 reaches 1,844 Elo in GDPval-AA v2.1, Opus 5.5 holds 1,846, the gap is minimal. In AA-Briefcase v1.1 it is 1,811 against 1,822. Humanity's Last Exam with tools: 64.5 against 67.7 percent. On computer use, OSWorld 2.1 in the partial evaluation, it is 80.1 against 57.0 percent for the predecessor, Opus 5.5 holds 81.8 there. Chart recognition without tools, Chartography: 61.6 against 15.6. the-decoder.com notes that Sonnet 5.5 is the first Sonnet model able to play through Pokémon Red using screenshots only. The point of this list is not the collection but the spread: The same model family covers both the multiplication of its predecessor's score and the two-point difference to the top model, depending on which task and which level is measured.

Sonnet 5.5 offers five effort levels: low, medium, high, xhigh and max. The defaults are split: In the Claude apps, Medium is preset, on the platform it is High. Per Anthropic, Sonnet 5.5 at Low or Medium beats Sonnet 5's best score on several benchmarks at roughly one tenth of the cost per task, and at High it reaches GPT-6 Sol's best score on FrontierCode 1.1 at about one fifth of the cost. The level choice is thus the actual price rule of this model, not the token price: same list, same task, same price table, and still the cost per task swings by a factor of twenty and more, exactly as in the pelican attempt. Whoever turns the level dial in the app operates a lever the token price does not advertise.

The independent evaluation site Artificial Analysis reports on 2026-09-30: Claude Sonnet 5.5 reaches 56 points on the Intelligence Index, rank 2, only two points behind Opus 5.5 at Max. At the same time the site measures the highest token consumption it has ever recorded: roughly 193,000 output tokens per task at the Max level, about 60 percent above Opus 5.5 at Max and roughly seven times GPT-6 Astra at Max. On cost it lists $7.60 per task, about half above Sonnet 5. The independent site reports exactly the pattern the pelican attempt showed: More effort buys mainly more tokens with this model, not proportionally more hits. The limit: Artificial Analysis measures with its own harness and task sample, their figures are not comparable one to one with Anthropic's, but the direction agrees in both sources.

From the effort jump follows the integration advice: Anyone embedding the model through the API should fix the effort level and measure cost per task, not per token. The token price is constant, task cost swings by a factor of twenty and more across effort levels, as the pelican attempt shows. A dashboard that only sums tokens hides exactly the span in which the model gets expensive. The second tip follows from the FrontierCode result: The default assumption that more effort brings better results does not hold, on FrontierCode 1.1 the curve turns down at Max. Whoever wants to pick a level should look for the number that measures their own task, not the peak score on somebody else's benchmark.

A second case shows the same lesson at a different spot: The company behind Holo 4 has released two agent models that can operate a computer. On the OSWorld 2.0 test, which grades long workflows on the desktop, the large variant with 27 billion parameters reaches 61.7 percent, the smaller one 30.9. The catch: Only the smaller variant is under the Apache 2.0 license and may run commercially in your own operation. The larger one may be used commercially only through the company's interface. Whoever chooses by benchmark ends up in the variant they may not self-host for client projects. Whoever chooses by license gets the variant they may host and accepts the lower score. The order given in the news item therefore reads: first sort out the license, then define where a human signs off before money moves or a message goes out, and only then compare percentages.

The third item turns the order around once more. In a study with email agents the picture is: The best tested configuration passes only 33.5 percent of scenarios, while 99.7 percent of the tool calls ran through without errors. A valid call is simply not a completed task. The gap between 99.7 and 33.5 is the part where agents fail, not at execution but at the goal: the question of whether the action the agent triggered is even the task that was meant. For operators this means: The sign-off chain before money flows or a message gets sent is not a formality, it is the actual place where errors land.

In my setup this translates into three concrete moves: Assign the effort level as a fixed setting, and measure task cost per task type, not model averages. License before benchmark: With Holo 4, the license text decides whether a model is hostable for client projects, the percentages come later. Sign-off chain before percentage comparison: The order from the news item, license, sign-off, then percentages, is the order in which I evaluate setups. The pelican test is the short form of all of this: More thinking effort is not more result, it is more cost with an unknown direction of outcome, and the lever that balances it is called the sign-off chain.

The limits, explicitly. The Sonnet 5.5 numbers come from Anthropic's announcement, independent re-measurements are available via Artificial Analysis, and the Anthropic figures are self-reported. The Holo 4 numbers come from the news item, and the name of the company behind Holo 4 is not cleanly recognizable in the automatic transcript, so this post deliberately keeps only the label 'the company', a check against the license text on the model pages remains necessary before hosting decisions. The email-agent study is not named with a DOI or preprint link in the item, that gap stays open here because no primary source is available. The pelican attempt is a single case, not a benchmark, it illustrates the effort jump but does not prove it.

I am Marcel, a graphic designer running the HUNTER cyberdeck on a Google Pixel 6a, with my own agent setup of Hermes, pi and herdr. All numbers in this post come from Anthropic's announcement and the AI news item of 2026-09-30, the video transcript is input, the Anthropic page is the source of record. Last fact-checked on 2026-09-30.

Sources: Introducing Claude Sonnet 5.5, Anthropic, retrieved 2026-09-30. Claude Sonnet 5.5 Benchmarks Explained, vellum.ai, retrieved 2026-09-30. Anthropic's Claude Sonnet 5.5 nearly matches Opus 5.5 on benchmarks, the-decoder.com, retrieved 2026-09-30. Video input: AI news item of 2026-09-30 (Sonnet 5.5, Holo 4, agent licenses), YouTube, channel Chilikanal, transcript retrieved 2026-09-30. English quotations are kept verbatim in the post and translated immediately.