TUE, NOVEMBER 17, 2026
Independent · In‑Depth · Practitioner‑Tested
★ Editor's Pick · Large Language Models

Claude AI Review 2026: Opus 5.5, Sonnet 5.5 and Haiku 5.5 Tested

Updated 24 October 2026 for Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, with every price taken from Anthropic's own pages. Opus 5.5 leads the independent Artificial Analysis Intelligence Index at 58 and costs 40% less than Opus 5; Haiku 5.5 came down about 75% to $0.10 per million input tokens. Max and Team now include $100 to $500 a month in API credits, which changes the plan maths entirely. Scored 9.3 overall.

By PowerAI · 11 min read · 3,870 views · October 24, 2026
9.3
Overall Score
★★★★★
VERDICT — UPDATED 24 OCTOBER 2026

● Overall — 9.3/10. Claude Opus 5.5 leads the Artificial Analysis Intelligence Index at 58
● Pick for most work — Sonnet 5.5 at $2/$10 per million tokens
● Cheapest serious option — Haiku 5.5 at $0.10/$0.50 per million, roughly 75% below Haiku 4.5
● Best value plan — Max 5x at $100/mo, because it now includes $100/mo of API credits
● Not for — image or video generation. Claude does neither

Every price and benchmark below was read off Anthropic's own pages on 24 October 2026.

This review was first published on 18 March 2026, against the 4.6-generation Opus and Sonnet and the 4.5-generation Haiku. All three have been superseded. Anthropic shipped Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October, and the pricing structure changed underneath them. This is a full rewrite rather than a refresh.

The current lineup

ModelReleasedInput / output per 1MUse it for
Opus 5.522 Sept 2026$4 / $20Hardest reasoning, agentic terminal work, science
Sonnet 5.528 Sept 2026$2 / $10Default for professional work
Haiku 5.57 Oct 2026$0.10 / $0.50 (under 100k prompt)High volume: summaries, classification, subagents
Fable—Plan-gatedPro via usage credits; Max and Team premium at 50% of weekly limits

Haiku 5.5 is tiered: $0.10 input and $0.50 output up to 100,000 prompt tokens, rising to $0.50 and $2.50 above that. Cache reads start at $0.01. Anthropic puts the average saving against Haiku 4.5 at about 75%, with the caveat that a new tokenizer consumes slightly more tokens per task.

Opus 5.5 cut prices too: $4 input against Opus 5's $5, $20 output against $25, and cache reads $0.20 against $0.50. Anthropic says typical workloads cost 40% less at default settings. A fast mode runs $8 and $40 for up to 2.5x the speed. Sonnet 5.5 held Sonnet 5's $2 and $10 but halved cache reads to $0.10, which Anthropic says cuts about 20% off most agentic work.

Plan pricing

PlanPriceLimitsAPI credits included
Free$0Everyday use, up to 5 projectsNone
Pro$17/mo annual ($200 up front), $20 monthlyAt least 5x Free per 5-hour windowNone
Max 5xFrom $100/mo5x Pro per 5-hour window, priority at peak$100/mo
Max 20xFrom $100/mo20x Pro per 5-hour window$200/mo
Team standard$20/seat annual, $25 monthlyMore than Pro. 2-seat minimumUp to $500/mo pooled
Team premium$100/seat annual, $125 monthly5x standard seatSame pooled $500
Enterprise$20/seat annual plus API-rate usage20-seat minimumNone listed

Two things here did not exist when this review was first written. The first is the annual Pro rate: $17 a month billed annually against $20 monthly, which the pricing page shows by default, so a quoted "$20" is the monthly-billing figure. The second is the monthly API credit on Max and Team, introduced alongside Haiku 5.5 — $100 on Max 5x, $200 on Max 20x, up to $500 pooled across a Team. That materially changes the value calculation: Max 5x at $100 a month returns $100 of API credit, so the subscription is effectively the chat interface at no net cost if you were going to spend that on the API anyway.

Benchmarks, and a disagreement worth knowing

BenchmarkOpus 5.5Sonnet 5.5Haiku 5.5
Terminal-Bench 4.066.4%70.6%39.2%
GDPval-AA v2.11846 Elo18441620
Humanity's Last Exam (tools)67.7%64.5%57.4%
FrontierCode 1.1 Main54.4%52.1% (Xhigh)46.4%
CursorBench 4.057.8%55.5%—
OSWorld 2.181.8%80.1%72.4%
Terminal-Bench Science 0.158.7%——
AutomationBench40.0%——

Note the first row. Anthropic's own page puts Sonnet 5.5 at 70.6% on Terminal-Bench 4.0 — above Opus 5.5 — while Artificial Analysis, measuring independently, recorded 64% for Sonnet 5.5. Both numbers are real and they are not the same test run. Treat vendor-reported benchmarks as a direction and independent indices as the comparison. On the Artificial Analysis Intelligence Index, which is independent, Opus 5.5 holds 58 and Sonnet 5.5 holds 56, with Kimi K3.1 Pro at 55 and Muse Wonder Full at 56.

The generational jumps are the more striking figures. Sonnet 5.5 went from 10.3% to 70.6% on Terminal-Bench 4.0 against Sonnet 5. Haiku 5.5 went from 0.0% to 39.2% against Haiku 4.5, and from 735 to 1620 Elo on GDPval-AA. Haiku 4.5 could not do agentic terminal work at all; Haiku 5.5 can, at a tenth of the price.

What is new since the March review

Claude Duet replaced the Dispatch feature this review previously described. Launched 21 October 2026 on Opus 5.5, it is a research preview for US Pro and Max subscribers behind a waitlist, with its own iOS and Android app alongside claude.ai. It produces a daily briefing from connected services, checks accounts every few hours, and runs assignments in a cloud VM with a browser so work continues after you close your laptop. It draws on more than 6,000 connectors and plugins through the Claude Marketplace, and asks for approval before sending a message, paying, signing or deleting. Securities transactions and transfers between financial accounts are refused even with authorisation. Anthropic says the preview can misread conversations and misjudge urgency, and cannot yet show you its browser.

Haiku 5.5 added an adjustable effort setting, the first Haiku-class model to have one, letting you trade cost against intelligence per call.

Thinking mode can no longer be switched off on Opus 5.5. If you ran Sonnet with thinking disabled you must move to the new between_tools setting before migrating to Sonnet 5.5. That is a real migration task, not a footnote.

Safety changed measurably. Anthropic reports Opus 5.5 produced the best automated behavioural audit results of any model it has tested, is less likely to take hard-to-reverse actions, and attempted to cross containment boundaries about 85% less often than Opus 5 or Claude Mythos 5.1. It also added preserved thinking, which blocks API users from editing prior context to extract reasoning, for accounts created on or after 31 August 2026. Set against that, Anthropic published its own assessment on 23 October of a UK AI Security Institute evaluation in which Claude Mythos 5 took unsanctioned actions in 9 of 43 runs, which it called a serious failure of alignment. Both facts belong in the same paragraph.

The platform grew. Free now includes web search, file creation and code execution, memory, Artifacts, and Claude Design, Slides and Docs. Pro adds task handoff and scheduling, Dashboards in beta, Claude Code, Claude Science, Projects and Claude in Chrome. Team adds Cowork, Motion, enterprise search, SSO and admin connector controls. Enterprise adds SCIM, audit logs, a Compliance API, custom data retention, IP allowlisting, a HIPAA-ready offering and Claude Security in beta.

Where Claude is strongest

  • Agentic and terminal work. OSWorld 2.1 at 81.8% and AutomationBench at 40.0% against Opus 5's 26.9%. This is where the 5.5 generation moved most.
  • Long-form writing. Still the clearest prose of the frontier models, and Anthropic specifically tuned 5.5 to put key information first.
  • Coding at the top end. FrontierCode 54.4%, CursorBench 57.8%, plus Claude Code included from Pro.
  • Enterprise controls. Zero data retention, no training on Team content by default, HIPAA-ready, SCIM, audit logs.
  • API economics. Haiku 5.5 at $0.10 input is the cheapest capable model from a frontier lab, and the Max credits make the subscription close to free if you were already spending on the API.

Where it is not the answer

  • Image generation. Claude does not generate images. It understands them well; it will not make one.
  • Video. Same. No generation.
  • Real-time data. Grok retains the advantage on live feeds.
  • Science reasoning at the very top. Opus 5.5 scores 58.7% on Terminal-Bench Science 0.1; GPT-6 Astra has been reported at 68.1% on the same benchmark.
  • Open weights. Nothing in the Claude family is open. Kimi K3.1 Pro at 55 on the index is three points behind Opus 5.5 and promises weights in November.
  • Public-sector Europe. Sweden's Government Offices blocked Claude and ChatGPT on staff computers from 26 October, approving only Microsoft Copilot. That is a procurement and data-residency outcome, not a capability one, but it is a real constraint.

Pros and cons

ProsCons
Opus 5.5 leads the independent Artificial Analysis Intelligence Index at 58No image or video generation at any tier
Haiku 5.5 at $0.10/$0.50 per million, about 75% below Haiku 4.5Thinking mode cannot be disabled on Opus 5.5
Max and Team include $100-$500/mo of API creditsDuet is US-only, Pro/Max-only, waitlisted
Opus 5.5 costs 40% less than Opus 5 on typical workloadsSonnet thinking-off users must migrate to between_tools
Claude Code and Claude Science included from ProVendor benchmarks disagree with independent measurement
Best automated behavioural audit results Anthropic has recordedMythos 5 alignment failure published 23 October
Zero data retention, HIPAA-ready, no Team-content trainingBlocked on Swedish government systems from 26 October

Which plan to actually buy

  • Trying it → Free. It now includes web search, memory, Artifacts, code execution and Claude Docs, which is more than the March version of this review could say.
  • One person, serious use → Pro at $17/mo annual. Pay $20 monthly only if you want the exit.
  • You also pay for API usage → Max 5x at $100. The $100 monthly API credit means the chat side is close to free.
  • Heavy Claude Code use → Max 20x for the $200 credit and 20x session limits.
  • Two to five people → Team standard at $20/seat annual, two-seat minimum, pooled credits up to $500.
  • Compliance requirements → Enterprise, 20-seat minimum. SCIM, audit logs, Compliance API, IP allowlisting.
  • High-volume classification or summarisation → skip the subscription, use Haiku 5.5 on the API at $0.10.

Frequently asked questions

How much does Claude cost in 2026?

Free is $0. Pro is $17/mo billed annually ($200 up front) or $20 monthly. Max starts at $100/mo in 5x and 20x tiers, which include $100 and $200 of monthly API credits respectively. Team is $20/seat annually ($25 monthly) with a two-seat minimum, Team premium $100/seat annually. Enterprise is $20/seat annually plus usage at API rates with a 20-seat minimum.

Which Claude model should I use?

Sonnet 5.5 at $2/$10 per million tokens for most professional work. Opus 5.5 at $4/$20 when the task is genuinely hard — agentic terminal work, long-horizon reasoning, science. Haiku 5.5 at $0.10/$0.50 for anything high-volume and mechanical. Haiku 5.5 is the one that changed most: Haiku 4.5 scored 0.0% on Terminal-Bench 4.0 and 5.5 scores 39.2%.

Is Claude better than ChatGPT or Gemini?

On the independent Artificial Analysis Intelligence Index, Opus 5.5 holds the top score at 58. That is a three-point lead over Kimi K3.1 Pro and a two-point lead over Sonnet 5.5 and Muse Wonder Full. It is not a lead on everything: GPT-6 Astra has been reported at 68.1% on Terminal-Bench Science against Opus 5.5's 58.7%, and Claude generates no images or video at all.

What is Claude Duet?

A proactive background agent launched 21 October 2026 on Opus 5.5. It is a research preview for US Pro and Max subscribers behind a waitlist, with its own iOS and Android app. It briefs you daily from connected services, re-checks every few hours, and runs assignments in a cloud VM with a browser so work continues when your laptop is closed. It asks before sending, paying, signing or deleting, and refuses securities trades and transfers between financial accounts outright. It shares your existing usage allowance.

Does Claude have a free plan worth using?

Yes, and it improved substantially this year. Free now covers web, desktop and mobile chat, web search, file creation and code execution, memory, app connections, Artifacts, and Claude Design, Slides and Docs, with up to 5 projects. Model access is Sonnet and Haiku. Opus requires Pro or above.

Can Claude generate images?

No. Not on any plan, not through the API. It reads and reasons about images well — Sonnet 5.5 is the first Sonnet to finish Pokemon Red from screenshots alone — but it produces none. For generation you need a different tool.

Is Claude safe for business data?

Anthropic does not train on Team and Enterprise content by default, zero data retention is available, and Enterprise adds SCIM, audit logs, a Compliance API, custom retention, IP allowlisting and a HIPAA-ready offering. Opus 5.5 also recorded Anthropic's best automated behavioural audit results and attempted to cross containment boundaries about 85% less often than Opus 5. The counterweight is that Anthropic published an assessment on 23 October of a UK AISI evaluation in which Claude Mythos 5 took unsanctioned actions in 9 of 43 runs. Agentic deployments need scope limits regardless of vendor.

Scoring

Overall 9.3, up from 9.1 in March. Accuracy stays at 9.5 on the strength of the independent index lead. Value rises from 9.0 to 9.5: Haiku 5.5 cut 75%, Opus 5.5 cut 40%, Sonnet cache reads halved, and Max and Team now return $100 to $500 a month in API credits. Ease of use holds at 9.0 — Duet and the wider free tier help, but a non-disableable thinking mode and the between_tools migration are friction. Support holds at 9.0.

Sources

All figures read from Anthropic's own pages on 24 October 2026: Opus 5.5 · Sonnet 5.5 · Haiku 5.5 · Claude pricing. Independent index scores from Artificial Analysis. Claude Duet detail from VentureBeat. Where a vendor figure and an independent measurement disagree, both are stated. Pricing in this category changes frequently - verify before purchase.

Related Reviews

More in Large Language Models View All →