AI Shqipe
valon-hajredini.aishqipe.com/posts/claude-sonnet-5-5-benchmarks-price
Valon Hajredini
Subscribe

Claude Sonnet 5.5: Opus-Level Scores at Half the Price, and Where It Still Falls Short

Anthropic shipped Sonnet 5.5 on 28 September 2026 at $2 in and $10 out per million tokens, the same as Sonnet 5. It beats Opus 5.5 on Terminal-Bench 4.0 and ties it on knowledge work. Here is what Anthropic's charts show, what Artificial Analysis found, and when you should still pay for Opus.

Anthropic released Claude Sonnet 5.5 on 28 September 2026, six days after Opus 5.5. The price per token did not move. What moved is everything else: Anthropic says it writes more than 30% faster than Sonnet 5, finishes tasks in fewer tokens, and on several of its own benchmarks lands within a point or two of Opus 5.5, a model that costs twice as much. On one, it beats it.

Flat illustration of a desk seen from above: a laptop with code, a spreadsheet, slides and a document, each linked by thin terracotta lines to a small glowing spark in the centre.
Sonnet 5.5 is pitched at the work Sonnet has always been for: well scoped, everyday and high volume.

Anthropic's own numbers

Anthropic compared Sonnet 5.5 against Sonnet 5 and Opus 5.5. The jump from Sonnet 5 is huge almost everywhere. The gap to Opus 5.5 is small almost everywhere, and on Terminal-Bench 4.0 it runs the other way.

Terminal-Bench 4.0 — Sonnet 5.5: 70.6%Terminal-Bench 4.0 — Opus 5.5: 66.4%Terminal-Bench 4.0 — Sonnet 5: 10.3%CursorBench 4.0 — Sonnet 5.5: 55.5%CursorBench 4.0 — Opus 5.5: 57.8%CursorBench 4.0 — Sonnet 5: 34.1%Humanity's Last Exam — Sonnet 5.5: 64.5%Humanity's Last Exam — Opus 5.5: 67.7%Humanity's Last Exam — Sonnet 5: 54.9%OSWorld 2.1 — Sonnet 5.5: 80.1%OSWorld 2.1 — Opus 5.5: 81.8%OSWorld 2.1 — Sonnet 5: 57%Chartography — Sonnet 5.5: 61.6%Chartography — Opus 5.5: 64.4%Chartography — Sonnet 5: 15.6%FrontierCode 1.1 — Sonnet 5.5: 52.1%FrontierCode 1.1 — Opus 5.5: 54.4%FrontierCode 1.1 — Sonnet 5: 42.4%0%20%40%60%80%100%Terminal-Benc…CursorBench 4…Humanity's La…OSWorld 2.1ChartographyFrontierCode…ScoreTerminal-Bench 4.0 — Sonnet 5.5: 70.6%Terminal-Bench 4.0 — Opus 5.5: 66.4%Terminal-Bench 4.0 — Sonnet 5: 10.3%CursorBench 4.0 — Sonnet 5.5: 55.5%CursorBench 4.0 — Opus 5.5: 57.8%CursorBench 4.0 — Sonnet 5: 34.1%Humanity's Last Exam — Sonnet 5.5: 64.5%Humanity's Last Exam — Opus 5.5: 67.7%Humanity's Last Exam — Sonnet 5: 54.9%OSWorld 2.1 — Sonnet 5.5: 80.1%OSWorld 2.1 — Opus 5.5: 81.8%OSWorld 2.1 — Sonnet 5: 57%Chartography — Sonnet 5.5: 61.6%Chartography — Opus 5.5: 64.4%Chartography — Sonnet 5: 15.6%FrontierCode 1.1 — Sonnet 5.5: 52.1%FrontierCode 1.1 — Opus 5.5: 54.4%FrontierCode 1.1 — Sonnet 5: 42.4%Terminal-Bench 4.0CursorBench 4.0Humanity's Last ExamOSWorld 2.1ChartographyFrontierCode 1.10%50%100%Score
  • Sonnet 5.5
  • Opus 5.5
  • Sonnet 5
Anthropic's published scores. Opus 5.5's Terminal-Bench figure is at xhigh effort; Humanity's Last Exam is with tools. Source: Anthropic's Sonnet 5.5 announcement.
Show the 18 numbers behind this chart
Anthropic's published scores. Opus 5.5's Terminal-Bench figure is at xhigh effort; Humanity's Last Exam is with tools. Source: Anthropic's Sonnet 5.5 announcement. — the numbers this chart is drawn from
CategorySonnet 5.5Opus 5.5Sonnet 5
Terminal-Bench 4.070.6%66.4%10.3%
CursorBench 4.055.5%57.8%34.1%
Humanity's Last Exam64.5%67.7%54.9%
OSWorld 2.180.1%81.8%57%
Chartography61.6%64.4%15.6%
FrontierCode 1.152.1%54.4%42.4%

The headline is Terminal-Bench 4.0, the agentic coding test run in a real shell. Sonnet 5 scored 10.3%. Sonnet 5.5 scores 70.6%, more than four points above Opus 5.5 at xhigh. A sevenfold jump in one release says as much about how badly Sonnet 5 handled long terminal sessions as about how good 5.5 is, but either way the problem is gone.

One oddity: on FrontierCode, Sonnet 5.5 scores 52.1% at xhigh but only 46.2% at max. Anthropic says max effort switches on a sub-agent code review pass that sometimes times out or makes changes outside the task. More effort is not always better, and here it is measurably worse.

The knowledge work benchmarks, scored as Elo ratings, are where Sonnet 5.5 gets closest to Opus. On GDPval-AA it is two points behind, which is a tie. Both numbers come from Artificial Analysis running a pre-release build with a structured output bug that has since been fixed, so they may move slightly.

GDPval-AA v2.1 — Sonnet 5.5: 1844GDPval-AA v2.1 — Opus 5.5: 1846GDPval-AA v2.1 — Sonnet 5: 1449AA-Briefcase v1.1 — Sonnet 5.5: 1811AA-Briefcase v1.1 — Opus 5.5: 1822AA-Briefcase v1.1 — Sonnet 5: 13590500100015002000GDPval-AA v2.1AA-Briefcase v1.1EloGDPval-AA v2.1 — Sonnet 5.5: 1844GDPval-AA v2.1 — Opus 5.5: 1846GDPval-AA v2.1 — Sonnet 5: 1449AA-Briefcase v1.1 — Sonnet 5.5: 1811AA-Briefcase v1.1 — Opus 5.5: 1822AA-Briefcase v1.1 — Sonnet 5: 1359GDPval-AA v2.1AA-Briefcase v1.1010002000Elo
  • Sonnet 5.5
  • Opus 5.5
  • Sonnet 5
GDPval-AA v2.1 and AA-Briefcase v1.1 Elo ratings. Source: Anthropic, run by Artificial Analysis on a pre-release build.
Show the 6 numbers behind this chart
GDPval-AA v2.1 and AA-Briefcase v1.1 Elo ratings. Source: Anthropic, run by Artificial Analysis on a pre-release build. — the numbers this chart is drawn from
CategorySonnet 5.5Opus 5.5Sonnet 5
GDPval-AA v2.1184418461449
AA-Briefcase v1.1181118221359

Faster, and fewer tokens per task

The price sheet is the same as Sonnet 5 apart from cheaper cache reads. The saving Anthropic is selling is efficiency: up to 30% less per task because the model gets there in fewer tokens and fewer steps, and it now batches tool calls instead of making them one at a time.

Two abstract streams of shapes on cream: a short, dense, fast-moving stream above and a long, loose, winding stream below.
Same price per token, fewer tokens per job: that is where the up to 30% saving comes from.

The early customer numbers in the launch are consistent with that. They are vendor-selected, so read them as direction rather than proof.

  • Balyasny Asset Management: about 121,000 tokens per answer across 2,441 finance tasks, against 497,000 for Sonnet 5.
  • Base44: 3.6 iterations per app build over 118 builds, against 7.7 for Opus 5, with the fewest failed tool calls of any model it tested.
  • Lovable: about a third fewer tool calls and half as many shell executions to finish a task.
  • Zendesk: tickets processed 20% faster, with fewer wrong escalations. Atlassian: Rovo agent actions up to 30% faster.

Anthropic also says that at low or medium effort Sonnet 5.5 beats Sonnet 5's best scores on Terminal-Bench, CursorBench and AA-Briefcase for roughly a tenth of the cost per task. If that holds up, the right move for most teams is not to switch and keep their settings, but to switch and turn effort down.

Artificial Analysis: second place, and a token bill at max

Artificial Analysis puts Sonnet 5.5 at 56 on its Intelligence Index at max effort. That is second on the board, two points behind Opus 5.5 at 58, three ahead of Fable 5.1, and 18 ahead of Sonnet 5.

Opus 5.5: 58Sonnet 5.5: 56Fable 5.1: 53Sonnet 5: 380204060Opus 5.5Sonnet 5.5Fable 5.1Sonnet 5Index scoreOpus 5.5: 58Sonnet 5.5: 56Fable 5.1: 53Sonnet 5: 38Opus 5.5Sonnet 5.5Fable 5.1Sonnet 50204060Index score
Artificial Analysis Intelligence Index, each model at its best effort setting.
Show the 4 numbers behind this chart
Artificial Analysis Intelligence Index, each model at its best effort setting. — the numbers this chart is drawn from
CategoryIndex score
Opus 5.558
Sonnet 5.556
Fable 5.153
Sonnet 538

The effort curve is where it gets interesting. Each step up buys about five index points, but the cost per task does not rise in step with it: from high to xhigh it more than doubles, and from xhigh to max it nearly triples again.

Low — Index score: 36Medium — Index score: 41High — Index score: 47Xhigh — Index score: 52Max — Index score: 56354045505560LowMediumHighXhighMaxIndex scoreEffortLow — Index score: 36Medium — Index score: 41High — Index score: 47Xhigh — Index score: 52Max — Index score: 5630405060LowHighMaxIndex scoreEffort
Sonnet 5.5 on the Artificial Analysis Intelligence Index by effort level.
Show the 5 numbers behind this chart
Sonnet 5.5 on the Artificial Analysis Intelligence Index by effort level. — the numbers this chart is drawn from
EffortIndex score
Low36
Medium41
High47
Xhigh52
Max56
Low: 0.41Medium: 0.59High: 1.08Xhigh: 2.74Max: 7.602468LowMediumHighXhighMaxUSD per taskEffortLow: 0.41Medium: 0.59High: 1.08Xhigh: 2.74Max: 7.6LowMediumHighXhighMax0510EffortUSD per task
Cost per Artificial Analysis index task in US dollars, by effort level.
Show the 5 numbers behind this chart
Cost per Artificial Analysis index task in US dollars, by effort level. — the numbers this chart is drawn from
EffortUSD per task
Low0.41
Medium0.59
High1.08
Xhigh2.74
Max7.6

At max, Sonnet 5.5 used about 193,000 output tokens per index task, the most Artificial Analysis has ever measured: around 60% more than Opus 5.5 at max and roughly seven times GPT-6 Astra. That makes a max effort task about 50% more expensive than on Sonnet 5, which cuts against the cheaper per task story. The two are not in conflict. The saving is real at low to high effort, and it disappears at max.

Where it beats Opus, and where it does not

Sonnet 5.5 ahead Terminal-Bench 4.0 (independent run): 64% vs 60%. AutomationBench-AA: 71% vs 70%. Hallucination rate on AA-Omniscience: 47% vs 59% (lower is better). Price: half of Opus 5.5 per token.
Opus 5.5 ahead Intelligence Index: 58 vs 56. AA-Omniscience factual accuracy: 66% vs 54%. Humanity's Last Exam and SciCode: about six points. CursorBench 4.0: 57.8% vs 55.5%.

The independent Terminal-Bench score is lower than Anthropic's, 64% against 70.6%, as it was for Opus 5.5. Different harnesses and time limits explain most of that, and Sonnet still comes out ahead. The 12 point gap on factual accuracy is the one to take seriously. Sonnet 5.5 hallucinates less often than Opus, but it gets fewer facts right. For research and fact heavy writing, Opus is still the safer pick.

Price

Sonnet 5.5 keeps Sonnet 5's $2 in and $10 out. Cache reads halve to $0.10 per million tokens, half of what Opus 5.5 charges. Cache writes are $2.50. Everything else is half the Opus price.

Input — Sonnet 5.5: 2Input — Opus 5.5: 4Output — Sonnet 5.5: 10Output — Opus 5.5: 20Cache write — Sonnet 5.5: 2.5Cache write — Opus 5.5: 5Cache read — Sonnet 5.5: 0.1Cache read — Opus 5.5: 0.205101520InputOutputCache writeCache readUSD per 1M tokensInput — Sonnet 5.5: 2Input — Opus 5.5: 4Output — Sonnet 5.5: 10Output — Opus 5.5: 20Cache write — Sonnet 5.5: 2.5Cache write — Opus 5.5: 5Cache read — Sonnet 5.5: 0.1Cache read — Opus 5.5: 0.2InputOutputCache writeCache read01020USD per 1M tokens
  • Sonnet 5.5
  • Opus 5.5
API price per million tokens. Source: Anthropic.
Show the 8 numbers behind this chart
API price per million tokens. Source: Anthropic. — the numbers this chart is drawn from
CategorySonnet 5.5Opus 5.5
Input24
Output1020
Cache write2.55
Cache read0.10.2

What else is new

  • The first Sonnet with cybersecurity safeguards. High risk cyber requests visibly fall back to Sonnet 5; routine work like finding and fixing bugs is not affected. Security professionals can apply to the Cyber Verification Program for tiered access.
  • The same anti distillation classifiers Anthropic uses on its top models, and zero data retention for eligible customers.
  • Available on the Claude apps, the Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Azure.
  • Claude Haiku 5.5 is due "in the coming weeks", aimed at high throughput, low cost work.

The verdict

For most everyday work, Sonnet 5.5 makes Opus 5.5 hard to justify. It matches it on knowledge work, beats it on agentic terminal work, and costs half as much per token. Two things keep Opus in the picture: it knows more, by a wide margin on factual recall, and it is the better choice for hard, judgement heavy tasks. And whichever you choose, do not leave Sonnet 5.5 on max. Its best value is at medium or high, where it already beats Sonnet 5's best for a fraction of the cost.

Sources: Anthropic, The Decoder, Vellum and Emergent for the Artificial Analysis results. For the bigger model, see Claude Opus 5.5: Every Benchmark.

Share

Related

Gemini 4 Argon: Google's New Frontier Model, the Benchmarks, the Price and What It Can Do
Claude Opus 5.5: Every Benchmark, From Anthropic, Cursor, Artificial Analysis, Sonar and ARC Prize
GPT-6 Astra and Claude Fable 5.1: The Benchmarks Side by Side