Claude Sonnet 5.5 review: GPT-6 Sol's price, and a Terminal-Bench score above Opus 5.5

Claude Sonnet 5.5, released by Anthropic on 28 September 2026, costs $2 per million input tokens, $10 per million output tokens and $0.20 per million for cache reads, the same list price as Claude Sonnet 5 and as OpenAI's GPT-6 Sol. Its 70.6% on Terminal-Bench 4.0 is Anthropic's own run at max effort, set against 66.4% for Claude Opus 5.5 at xhigh, Opus 5.5's best setting. Artificial Analysis's own run, both models at max, gives 63.6% and 59.6%.

Anthropic still writes that Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment."

The price per token hides the price per task. On both outside evaluators with results, Sonnet 5.5 at max costs more per task than Opus 5.5 at max: $7.60 against $5.98 per Artificial Analysis index task, and $20.78 against $6.19 per Cognition FrontierCode rollout. At high effort, the API default, it scores level with GPT-6 Sol's best on both.

No prompt was run through Claude Sonnet 5.5 for this review. The figures come from documents read on 28 September 2026: Anthropic's announcement, 148-page system card, pricing page and models overview; eight outside evaluators' pages, two with Sonnet 5.5 results; and 14 launch-day reports from Hacker News and GitHub.

Does Claude Sonnet 5.5 beat Opus 5.5 on Terminal-Bench?

Yes in both runs published so far, by 4.2 points in Anthropic's and 4.0 in Artificial Analysis's, but only Artificial Analysis ran the two models at the same effort. Anthropic's system card has Sonnet 5.5 at 70.6% at max and Opus 5.5 at 66.4% at xhigh, adding that for Opus 5.5 "at max effort it scores 64.8%, within noise of xhigh" (p.113).

Model	Effort	Harness	Score	Safeguard fallback in the run	Run by
Claude Sonnet 5.5	max	Claude Code, bare mode	70.6% (±2.5)	1.2% of requests, in 1.5% of trials	Anthropic
Claude Opus 5.5	xhigh	Claude Code, bare mode	66.4% (±2.6)	2.5% of requests, in 10% of trials	Anthropic
Claude Sonnet 5.5	max	mini-swe-agent	63.6%	default fallback on	Artificial Analysis
Claude Opus 5.5	max	mini-swe-agent	59.6%	default fallback on	Artificial Analysis

Claude Sonnet 5.5 System Card, pp.112-113; Artificial Analysis Intelligence Index v4.3.2 model pages, from each page's chart data. Read 28 September 2026. The ± values are the card's standard errors.

Anthropic's run covers 66 tasks, five trials each, with internet access blocked. The card claims no significance for the 4.2-point gap; by our arithmetic it is about 1.2 times the standard error of the difference, and the gap to Opus 5.5 at max is 5.8 points.

The benchmark's own leaderboard at tbench.ai had 15 entries on 28 September and none for Sonnet 5.5, Opus 5.5 or GPT-6 Sol, and the card notes that "OpenAI has not reported a Terminal-Bench 4.0 score for GPT-6 Sol, nor has the public leaderboard". Artificial Analysis's run, in a different harness, is the only outside check: it keeps the order at equal effort but sits 7 points below Anthropic's 70.6%.

Do fallback rates explain the gap?

The documents cannot say: the safeguard fallback touched 10% of Opus 5.5's trials and 1.5% of Sonnet 5.5's, but the card reports no score for the touched trials. abejora on Hacker News, 28 September 2026 cited the same figures and concluded that "just the difference in fall backs could probably explain the gap." That is the commenter's reading; the card neither makes the link nor rules it out.

Two details narrow it. The fallback model answered flagged requests, 2.5% of Opus 5.5's, not whole trials, so a touched trial was not necessarily a failed one. And Artificial Analysis ran both models with Anthropic's default fallback on and still has Sonnet 5.5 ahead by 4.0 points at the same effort.

Is Claude Sonnet 5.5 really up to 30% cheaper than Sonnet 5?

At high effort and below, yes: on Artificial Analysis's Intelligence Index v4.3.2, Sonnet 5.5's cost per index task is 40% below Sonnet 5's at high, 41% below at medium and 19% below at low (our arithmetic from its figures). At max it is 49% above, $7.60 against $5.09; on Artificial Analysis's cost breakdown about $1.76 of the $2.51 difference is input and cache tokens and $0.75 is output, from 192,800 output tokens per index task against Sonnet 5's 117,800 (our arithmetic).

Anthropic hedges the claim twice: Sonnet 5.5 "costs up to 30% less for most work", and "In our testing, it costs up to 30% less per task than its predecessor." At xhigh, the saving on Artificial Analysis's figures is 5% (our arithmetic). The system card carries neither the 30% figure nor the speed claim, and twice ties a result in Anthropic's alignment audit to Sonnet 5.5 writing longer outputs (pp.57 and 84). Speed cannot be checked yet: Artificial Analysis shows "Speed N/A" on every Sonnet 5.5 entry.

Launch-day readers were cool on the saving. onlyrealcuzzo on Hacker News, 28 September 2026 quoted the 30% line and answered: "This isn't enough. Sonnet 5 was arguably the most cost ineffective model ever released at the time of a release." ghoshbishakh, 28 September 2026 asked whether Sonnet 5.5 at max "is as expensive as Fable 5.1? Because it uses a ton of tokens for a task." No Fable 5.1 cost was read for this review, but Cognition's page answers the token half: $20.78 per FrontierCode rollout at max, 3.4 times Opus 5.5 at max.

What effort does Claude Sonnet 5.5 use by default?

High on the Claude Platform and API, medium in Claude Code and Anthropic's apps. On Artificial Analysis's Intelligence Index v4.3.2 those two defaults score 47 at $1.08 per index task and 41 at $0.59; max scores 56 at $7.60.

The split is in the announcement: "In Claude Code and our apps, the default effort is set to Medium, while the Claude Platform defaults to High." The models overview lists high for Sonnet 5.5 and medium for Opus 5.5, so on the API the cheaper model starts one setting higher. The system card states no default; its summary table uses "adaptive thinking at max effort, default sampling settings (temperature, top_p), averaged over 5 trials".

Effort	Intelligence Index v4.3.2	Cost per index task	Terminal-Bench 4.0 (Artificial Analysis)	FrontierCode 1.1 Main (Cognition)	Cost per FrontierCode rollout
low	36	$0.41	20.7%	29.3%	$0.19
medium, Claude Code and apps default	41	$0.59	29.8%	36.5%	$0.24
high, Platform default	47	$1.08	43.9%	49.4%	$0.42
xhigh	52	$2.74	57.1%	52.1%	$1.59
max	56	$7.60	63.6%	46.2%	$20.78

Artificial Analysis model pages, each labelled "Default Fallback"; Cognition FrontierCode 1.1, claude-code harness. Read 28 September 2026.

Thinking is "Adaptive" for Sonnet 5.5 in the models overview, where Opus 5.5 is "Adaptive (always on)". Developers who ran Sonnet 5 with thinking off must move to a new between_tools setting, which the announcement says "keeps up-front thinking off". Artificial Analysis has no reasoning-off entry for Sonnet 5.5.

Why is the default medium in the apps and high on the Platform?

Anthropic does not say. Its only stated reasoning is general: "At lower settings, Claude answers faster and uses fewer tokens, which suits routine work." On the figures above, the step from medium to high adds 6 index points for 49 cents per index task, 14.1 points on Artificial Analysis's Terminal-Bench run, and 12.9 points on FrontierCode for 18 cents a rollout.

Why does Claude Sonnet 5.5 score lower at max effort than at xhigh on FrontierCode?

Anthropic's announcement gives the reason in its footnote 2: at max, Sonnet 5.5 more often ran Claude Code's code-review skill, which splits a review across many subagents, and in two cases Cognition examined this led to a timeout or to edits beyond the task's scope. The scores are 46.2% at max and 52.1% at xhigh on FrontierCode 1.1 Main, and Cognition's page shows both.

The system card does not carry that mechanism. Its note is that the evaluation "targets mergeable code diffs that need no human edits. Its grading penalizes out-of-scope changes, even if they are high quality or helpful" (p.111). It reports the same inversion on the Extended set, 64.4% at xhigh and 59.1% at max. Two cases out of 150 tasks are an illustration, not a count: neither document says how many max-effort runs timed out or strayed.

Cognition's page prices the inversion: at max, $20.78 per rollout and 595,700 output tokens; at xhigh, $1.59 and 47,400, so 13 times the cost for 5.9 points less (our arithmetic). The card plots tokens, not dollars, because "Cognition's result files report token counts for every model and no cost for Claude Sonnet 5.5"; Cognition's live page now shows costs, used here, and a dash for Sonnet 5.5's pass and flag rates.

The announcement's other FrontierCode claim, that at high effort Sonnet 5.5 "matches GPT-6 Sol's best score for about a fifth of the cost per task", holds on Cognition's page: 49.4% for $0.42 against Sol's 49.3% at max for $2.07. It is not in the card.

Why did Sonnet 5 score about 10% on Terminal-Bench and AutomationBench?

Neither of Anthropic's documents says. Sonnet 5's 10.3% on Terminal-Bench 4.0 appears only in the announcement's table, with no effort or harness named; the system card reports no Terminal-Bench score for Sonnet 5 at all. Its 10.7% on AutomationBench is a single cell in the card's Table 8.1.A, never mentioned in the AutomationBench section.

The low scores hold up outside Anthropic, and nobody explains them there either. The Terminal-Bench leaderboard lists Sonnet 5 at max in Claude Code at 12.4% ± 3.1%; Artificial Analysis has 14.1% at max; Zapier's AutomationBench 1.0.6 has "Claude Sonnet 5 (Max)" at 10.65% for $0.88 a task. nicoburns on Hacker News, 28 September 2026 put it plainly: "5 was definitely bad."

So the "10.3% -> 70.6%" that pavitheran on Hacker News, 28 September 2026 quoted in the same minute, calling it "impressive especially for the cost", starts from a baseline the system card does not report.

What do outside evaluators say about Claude Sonnet 5.5's scores?

Two had published by 18:41 UTC on launch day. Artificial Analysis scores Sonnet 5.5 56 on its Intelligence Index v4.3.2 at max effort, third of 216 models in its price band, and Cognition's FrontierCode 1.1 has 52.1% at xhigh as its best setting. Zapier's AutomationBench leaderboard, the Terminal-Bench 4.0 leaderboard, ARC Prize, LMArena, Epoch AI and METR had nothing.

Artificial Analysis labels every Sonnet 5.5 entry "Default Fallback", so its scores include answers from Anthropic's fallback model.

Which of Anthropic's Claude Sonnet 5.5 figures can you check?

Four of the eleven rows in the system card's summary table come from outside operators, and three check out: Cognition's FrontierCode figures match its page, and Artificial Analysis's GDPval-AA and AA-Briefcase figures match its data because they are the same measurement. Zapier's leaderboard has no Sonnet 5.5 row yet, so the 44.7% AutomationBench score rests on Anthropic's reporting. The other rows are Anthropic's own.

Evaluation, Table 8.1.A	Sonnet 5.5	Sonnet 5	Opus 5.5	GPT-6 Sol	Run by, and checkable
SWE-Bench Pro	81.3	63.2	89.9	not reported	Anthropic
FrontierCode v1.1 Main	46.2	42.4	54.4	49.3	Cognition; matches its page
OSWorld 2.1, partial score	80.1	57.0	81.8	not reported	Anthropic
HealthBench Professional, length-adjusted	69.2	57.8	65.6	not reported	Anthropic
GDPval-AA v2.1, Elo	1844	1449	1846	1487	Artificial Analysis, pre-release build; matches
AA-Briefcase v1.1, Elo	1811	1359	1822	1483	Artificial Analysis, pre-release build; matches
AutomationBench	44.7	10.7	42.5	32.0	Zapier, as reported by Anthropic; no row on Zapier's page

Seven of the eleven rows of Table 8.1.A, Claude Sonnet 5.5 System Card, p.109, max effort averaged over five trials; outside rows checked on 28 September 2026.

The GDPval-AA and AA-Briefcase numbers are the table's source, not a confirmation of it. Artificial Analysis ran them on a pre-release deployment that Anthropic later found had a bug affecting structured outputs, and Anthropic writes: "We expect the effect on Sonnet 5.5's scores, if any, to be small and to understate its performance." Artificial Analysis's pages do not mention the bug.

Three more rows need a note. OSWorld 2.1's 80.1 is the partial score; on the strict pass rate Sonnet 5.5 has 43.5% against Opus 5.5's 48.7%, a 5.2-point gap where the table shows 1.7 (p.130). Opus 5.5's 42.5 on AutomationBench is after its refused tasks were re-run with fallbacks; its own card reported 40.0 (p.137). And Sonnet 5.5's AutomationBench cost per task is "Anthropic's calculation from Zapier's recorded token counts", where every other model's cost is Zapier's published figure.

Is Claude Sonnet 5.5 safe?

Anthropic judges its misalignment risk low and finds it "broadly less capable than Opus 5.5 across domains", crossing no new threshold under its Responsible Scaling Policy (p.2). The same card lists regressions against Sonnet 5, among them a lower refusal rate for malicious requests in Claude Code, 85.2% against 87.9% (p.48).

The assessment is narrower than Opus 5.5's: only automated chemical and biological evaluations (p.12) and a targeted alignment review, because Sonnet 5.5 "does not advance the frontier of our models' capabilities" (p.55). A review of that section by Claude Mythos 5.1, printed in the card, adds that "the section has no dedicated reward-hacking evaluation of the kind the Claude Opus 5.5 card included" (pp.57-58).

Prompt injection is where the card reports the clearest gains: on Gray Swan's indirect prompt-injection benchmark, attacks succeeded 3.4% of the time at k=15, against 6.7% for Sonnet 5 (p.50), and in browser use "Without safeguards, Sonnet 5.5 is the first model we have evaluated with no successful attacks." (p.54). The self-preference test finds "a small but statistically significant bias toward itself" of 0.16 points on a 0 to 9 scale when the model is reminded it is Claude, slightly larger than Opus 5.5's (p.78).

How often did Claude Sonnet 5.5 try to escape its sandbox?

In 9 of 196 runs, 4.6%, in Anthropic's dedicated evaluation of 28 handmade scenarios run seven times each on isolated test machines: "well below earlier Claude models and below Claude Opus 5.5" (p.77). Opus 5.5's rate appears only in a chart.

The card is more favourable than the launch post. The announcement says Sonnet 5.5 "comes close to Opus 5.5, the best model we tested", while the card, in its dedicated evaluation, places it below Opus 5.5 and reports "the lowest rate of attempted sandbox escapes in our dedicated evaluation" (p.3).

Where did Claude Sonnet 5.5 regress against Sonnet 5?

In refusing misuse, in some harmlessness areas and on honesty. The largest measured drop is on malicious computer-use tasks, refused 79.46% of the time against Sonnet 5's 84.68%, the same rate as Opus 5.5 (pp.48-49).

Measure	Sonnet 5	Sonnet 5.5	System card
Malicious requests refused, Claude Code	87.9%	85.2%	Table 5.1.1.A, p.48
Malicious computer-use tasks refused	84.68%	79.46%	Table 5.1.2.A, p.48
Harmless responses, single-turn, API	96.65%	95.61%	p.33
Harmless responses, disordered eating, API	97.10%	95.29%	p.40

Higher is better in every row. Claude Sonnet 5.5 System Card, read 28 September 2026.

Other regressions have no single figure. In multi-turn testing Sonnet 5.5 "regressed in tracking and surveillance, violent extremism, and hate and discrimination on both the API and claude.ai" (p.32), and child-safety testing found "increased willingness to accept benign framing at face value" (p.37). It is "slightly more likely to state an incorrect answer than Sonnet 5" (p.80), has a lower MASK honesty rate (p.81), does not improve on Sonnet 5 on evasiveness, where the card finds the two similar and Sonnet 5.5 deflecting sensitive questions slightly more readily than most recent models (p.65), and its reasoning text is "the least legible of the models we tested" (p.72).

What happens when Claude Sonnet 5.5 falls back to Sonnet 5?

Claude Sonnet 5 answers instead. A request blocked by Sonnet 5.5's cyber classifiers falls back to Sonnet 5 automatically in Anthropic's own apps, and on the API only for developers who opt in (p.29). Blocks for chemical and biological misuse, conventional weapons and distillation have no fallback model (pp.10-11).

Sonnet 5.5 is the first Sonnet with these cyber safeguards, in Opus 5.5's three stages: a probe on internal activations, a lightweight classifier on Sonnet 5.5 itself, and a separate classifier model that decides with the probe (pp.28-29). Vulnerability discovery is allowed in source code and blocked in compiled binaries. The card warns that "users should expect increased refusals with Sonnet 5.5, even on benign cybersecurity-related tasks" (p.28); the announcement leads with reassurance: "Users can still find and fix bugs in their code as part of routine software development, but higher-risk cybersecurity tasks will visibly fall back to Sonnet 5."

In one of the card's tests the fallback model is the weak link. In a coding prompt-injection evaluation, "25% of requests sent to Sonnet 5.5 were served by Sonnet 5 after a cyber block, and 12.01% of those requests to Sonnet 5 were compromised", against 4 of the 5,901 requests Sonnet 5.5 answered itself (pp.51-52). On Gray Swan's benchmark, 11% of rollouts fell back and the attack rate did not rise (p.50).

The biology safeguards are described two ways. The announcement says Sonnet 5.5 "uses the same set of biology safeguards as Sonnet 5"; the card says "the same harmful CB misuse classifiers as for our deployment of Claude Opus 5" (p.10). Both can be true at once.

Why are Cyber Verification Program members still blocked?

Because Sonnet 5.5 is not in the program yet: the card says "Claude Sonnet 5.5 will be available through this program in the near future" (p.29), and the announcement says cyberdefenders will soon be able to apply to an expanded version.

Three launch-day reports came from members. johnmlussier on Hacker News, 28 September 2026 wrote: "Paying $200 a month and part of their Cyber Verification Program but can't use Opus 5.5 or Sonnet 5.5 for any authorized bounty work." John-Lussier, in a claude-code issue filed 28 September 2026 makes the same complaint: "Part of CVP but can't use latest models". AshamedBadger56 on Hacker News, 28 September 2026, whose employer joined, found that "it literally doesn't do anything or have a point" and that Claude "drops back to 4.8", a model name the card does not give.

Claude Sonnet 5.5 vs Claude Opus 5.5

Opus 5.5 costs twice as much per token, $4 and $20 against $2 and $10, yet on both outside evaluators Sonnet 5.5 at max costs more per task than Opus 5.5 at max, and Opus 5.5 at a lower effort scores about the same or better than Sonnet 5.5 at xhigh or max for about half the cost or less. Sonnet 5.5 is the cheaper route only at high effort and below.

Pairing	Claude Sonnet 5.5	Claude Opus 5.5	Publisher
Both at max	55.98, $7.60	57.62, $5.98	Artificial Analysis Intelligence Index v4.3.2, per index task
Sonnet max, Opus xhigh	55.98, $7.60	55.99, $3.46	same
Sonnet xhigh, Opus medium	51.85, $2.74	51.24, $1.34	same
Each at its best effort	xhigh, 52.1%, $1.59	medium, 54.6%, $0.80	Cognition FrontierCode 1.1 Main, per rollout

Artificial Analysis model pages and Cognition's FrontierCode page, read 28 September 2026.

Lower down the sources disagree: on Artificial Analysis's index, Opus 5.5 at low (42.31, $0.55) beats Sonnet 5.5 at medium, the Claude Code default (40.74, $0.59); on FrontierCode, Sonnet 5.5 at high (49.4%, $0.42) beats Opus 5.5 at low (47.3%, $0.40). Cache writes cost $2.50 on Sonnet 5.5 and $5.00 on Opus 5.5; reads cost $0.20 on both. In Anthropic's own table Sonnet 5.5 leads on two of eleven rows, AutomationBench (44.7 against 42.5) and length-adjusted HealthBench Professional (69.2 against 65.6); before adjustment the two tie at 77.1% (p.138).

Launch-day readers drew the same line. wkcheng on Hacker News, 28 September 2026 asked: "Why would you use Sonnet 5.5 on xhigh if you would get better results (higher score, cheaper cost) on Opus 5.5 high?" Both outside evaluators agree: Opus 5.5 at high scores 53.58 for $1.82 on Artificial Analysis and 54.0% for $1.09 on FrontierCode, against Sonnet 5.5 at xhigh on 51.85 for $2.74 and 52.1% for $1.59. square_usual, 28 September 2026 suggested that "the main point of this release is that you have a lower end than Opus low". schwarzrules, 28 September 2026, after a session limit on Opus 5.5, saw a use: "performance benefit of Sonnet 5.5 is a compelling alternative for managing session limits."

When does Anthropic still recommend Opus 5.5?

For hard, open-ended work, and as the starting point. The announcement says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment" and pitches Sonnet 5.5 as "strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets." The models overview tells unsure developers to "start with Claude Opus 5.5 for most workloads". The system card's nearest line: Sonnet 5.5 "is less capable than Claude Opus 5.5 on most evaluations we report" (p.12).

One customer quoted by Anthropic, Creator's Kevin Ngo, puts the split this way: "When Claude Opus 5.5 sets the architecture and general framework for a game, I would feel confident in letting Sonnet 5.5 implement it." The case for the larger model is in our review of Claude Opus 5.5.

Claude Sonnet 5.5 vs GPT-6 Sol

The two cost the same per token on all four lines, $2 input, $10 output, $0.20 cached input and $2.50 cache writes, and Sonnet 5.5 scores higher on both outside boards that list them together: 56 against 48 on Artificial Analysis's Intelligence Index at max, and 52.1% against 49.3% on Cognition's FrontierCode at each model's best effort. At max, Sonnet 5.5 costs 7.2 times as much per index task, $7.60 against $1.06.

At matched cost the gap closes. Sonnet 5.5 at high scores 46.74 for $1.08 per index task, level with Sol at max on 47.53 for $1.06, and on FrontierCode reaches 49.4% for $0.42 against Sol's best of 49.3% for $2.07. At max against max, Sonnet 5.5 leads Artificial Analysis's Terminal-Bench run 63.6% to 43.9%, but at matched cost that gap closes too: Sonnet 5.5 at high also scores 43.9%. Its lower hallucination rate, 47.0% against Sol's 60.1%, holds only at max; at high it is 64.6%.

Anthropic's table puts Sonnet 5.5 at 44.7 on AutomationBench against Sol's 32.0, Sol's max row on Zapier; Sol's best row there is xhigh, 33.2% for $0.27 a task. The announcement's footnote 4 adds that OpenAI "recently fixed a bug that degraded image understanding in GPT-6 Sol" and that Artificial Analysis's scores may not reflect the fix.

GPT-6 Sol reprices a whole request above 272,000 input tokens, as set out in our review of GPT-6 Sol; the Claude pricing page shows no such surcharge for Sonnet 5.5. Sol defaults to medium effort, Sonnet 5.5 on the API to high. No launch-day user report compared the two directly. How Opus 5.5 and Sol compare is in the Opus 5.5 and GPT-6 Sol head-to-head.

What do users say about Claude Sonnet 5.5?

Launch-day reports split between cost doubts and cyber blocks: of the 14 cited on this page, five question the saving, and three describe cyber safeguards blocking program members. All 14 were posted on 28 September 2026 between 17:59 and 18:33 UTC, 13 in Hacker News's launch thread and one on the anthropics/claude-code tracker. Reddit's launch-day comments had not reached the public archive, and G2, Trustpilot and Capterra list nothing for a model released that day. A first-hour sample records early opinion, not the model's quality.

Four reports rank the model: two put it level with or above Opus 5.5 on the benchmark numbers, one puts it above Sonnet 5 but below Fable, and one dismisses Sonnet models altogether. On launch day ramish94 found it "stacks up nearly 1:1 with Opus 5.5" on agentic coding benchmarks: "Sonnet 5.5 matches it and exceeds in some benchmarks." The commenter who called Sonnet 5 "definitely bad" added that 5.5 "seems a lot better so far. But still not close to Fable in terms of quality." jchw, 28 September 2026 dismissed the line: "Sonnet is a waste of time that does a bad job at a bad price." The fourth is the "10.3% -> 70.6%" comment above. And Leary, 28 September 2026 replied to the Terminal-Bench numbers: "And Sonnet 5.5 is more expensive than Opus 5.5 to hit that score on terminal bench!" No document read gives a Terminal-Bench cost to check that against.

Did Claude Sonnet 5.5 live up to Anthropic's launch claims?

On price per token, yes, and on cost per task at high effort and below; at max, no, and the speed claim has no outside figure yet.

What Anthropic said	What the documents and launch day showed
"costs up to 30% less for most work"	On Artificial Analysis's cost per index task against Sonnet 5 (our arithmetic): 40% less at high, 41% at medium, 19% at low, 5% at xhigh, 49% more at max
"runs 30%+ faster"	No outside speed figure yet, and not in the system card
Sandbox escapes: "comes close to Opus 5.5"	The card: 9 of 196 runs, below Opus 5.5
Routine software development unaffected by cyber safeguards	The card warns of more refusals "even on benign cybersecurity-related tasks"; three program members reported blocks on launch day

What don't we know yet about Claude Sonnet 5.5?

Its speed, and any result from Zapier, the Terminal-Bench leaderboard, ARC Prize, LMArena, Epoch AI or METR, none of which had published on Sonnet 5.5 when checked between 18:34 and 18:41 UTC on 28 September 2026. On launch day itself, Cognition's FrontierCode changelog recorded "Added Claude Sonnet 5.5.", with costs at every effort and no pass or flag rate, and Artificial Analysis listed five Sonnet 5.5 entries, all with default fallback and no speed.

Not yet published: Zapier's own row and cost; a Terminal-Bench leaderboard entry for Sonnet 5.5, Opus 5.5 or GPT-6 Sol; an Artificial Analysis launch article; Sonnet 5.5's place in the Cyber Verification Program, promised "in the near future"; and Claude Haiku 5.5, due "in the coming weeks". This page is re-checked against Anthropic's documents and every evaluator named here at least once a month, and whenever one of them publishes on Sonnet 5.5; each change will carry its date.