Claude Review 2026: Features, Pricing & Verdict

Which Claude plan to buy

Plan	Monthly	Annual (per month)	Seats
Free	$0	—	1
Pro	$20	$17 ($200 up front)	1
Max 5x	$100	—	1
Max 20x	$200	—	1
Team Standard	$25/seat	$20/seat	2–150
Team Premium	$125/seat	$100/seat	2–150
Enterprise	$20/seat + usage at API rates	billed annually	no published minimum

Prices verified 11 October 2026 against claude.com/pricing and Anthropic’s Max plan documentation. Our pricing tracker carries the dated change log.

Recommendation: Pro. Free gives you Sonnet and Haiku with caps that frustrate a real user inside about fifteen messages — fine for trying the product, not for working. Pro adds Opus, Fable on usage credits, Projects, the document and slide surfaces, Claude Code at entry limits, and the Chrome and Microsoft 365 extensions. At the same $20 as ChatGPT Plus it is the better buy for anyone whose work is mostly writing or code.

The one rung that is not priced linearly

Anthropic sells usage in multiples of a Pro plan, which makes the ladder easy to check. Divide the price by the multiple and almost every rung costs the same:

Rung	Price	Usage	Cost per 1x of Pro usage
Pro	$20	1x	$20
Max 5x	$100	5x	$20
Max 20x	$200	20x	$10
Team Standard	$25/seat	1.25x	$20
Team Premium	$125/seat	6.25x	$20

Max 20x is the only discount on the ladder: it buys twice the capacity per dollar that every other rung does. Going from Pro to Max 5x buys more headroom at exactly the same rate — so it is a decision about hitting caps, never about value. Going from Max 5x to Max 20x doubles the rate you get. Anthropic states the Team multiples directly: a Standard seat carries 1.25x the Pro plan’s per-session allowance and a Premium seat 6.25x, which works out to the same $20 per 1x as Pro and Max 5x — so a Premium seat is capacity, not a better rate.

One thing this table deliberately leaves out, because it is a different budget: every paid plan except Pro also returns a monthly API credit worth 1.00x the plan price at list — $100 on Max 5x, $200 on Max 20x, $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. The credits expire each billing cycle and Anthropic states they “don’t change your usage limits in Claude, Claude Code, or Claude Cowork,” so they buy API calls rather than chat or agent capacity. For a buyer who already has API spend they change the answer above; the full arithmetic is on the Claude Code review.

The same arithmetic now cuts the other way at OpenAI. From 30 October 2026 every rung of ChatGPT’s Pro ladder costs $20 per 1x of Plus usage, because Pro 200’s included usage halves while its price holds. That leaves Claude Max 20x as the only volume discount on offer among the big three assistant subscriptions. One honest caveat: “1x” is each vendor’s own multiplier against its own entry plan, and a Pro-unit of Claude is not the same quantity of work as a Plus-unit of ChatGPT. The table compares the shape of each ladder, not capacity across vendors.

Team has a two-seat minimum and a 150-seat ceiling, and annual billing takes 20% off both seat types. Enterprise is not quoted as “custom”: it is $20 per seat annually plus usage metered at API rates, which makes it the only Claude plan whose monthly cost you cannot predict from the plan name.

The model lineup (October 2026)

Model	Input	Output	Cache read	Context	Status
Claude Fable 5.1	$10	$50	$0.25	1M	ceiling
Claude Opus 5.5	$4	$20	$0.20	1M	flagship
Claude Sonnet 5.5	$2	$10	$0.10	1M	default
Claude Haiku 5.5 (prompt ≤100K)	$0.10	$0.50	$0.01	1M	new, 7 Oct 2026
Claude Haiku 5.5 (prompt >100K)	$0.50	$2.50	$0.05	1M	new, 7 Oct 2026
Claude Haiku 4.5	$1	$5	$0.10	200K	legacy
Claude Sonnet 4.5	$3	$15	$0.30	200K	retires 30 Nov 2026

Per million tokens, verified 11 October 2026 against Anthropic’s pricing documentation. Batch processing takes 50% off input and output on every model. Opus 5, Opus 4.8 and Sonnet 5 remain available at $5/$25 and $2/$10 respectively; Sonnet 5 is legacy with retirement no sooner than 30 June 2027.

Opus 5.5 is the flagship, at $4/$20 against Opus 5’s $5/$25 and with cache reads cut 60% to $0.20. Anthropic says it performs at Fable 5.1’s level on most work while costing 40% less to run — measured at default settings, and the default effort level dropped from high to medium in the same release, so roughly half of that saving is a configuration change rather than a price cut. Two things to plan around: flagged work does not run on the model you selected (cybersecurity requests fall back to Opus 4.8, biology and frontier-LLM requests to Opus 5, both billed at $5/$25 — more per token than Opus 5.5), and thinking can no longer be switched off. An optional Fast Mode runs at $8/$40, in research preview and on the first-party Claude API only. From 1 December 2026 Opus 5.5’s $4/$20 is matched to the decimal by Gemini 4 Argon, which scores five points lower on the one independent index and publishes no retirement date.

Sonnet 5.5 is the right default for most work: a June 2026 knowledge cutoff, a 512-token minimum cacheable prompt, and per-message effort control. Its cache read was cut from $0.20 to $0.10 on 7 October 2026, which is the one line of its rate card that no longer matches Sonnet 5 — the launch claim of complete price parity has lapsed in the buyer’s favour. Migrating from Sonnet 5 is a one-line model swap plus a six-item checklist; read the breaking changes before shipping, particularly the one that returns HTTP 200 while silently emptying the text an agent streams between tool calls.

Haiku 5.5 launched on 7 October 2026 and is the first Claude model priced by prompt length. Under 100,000 tokens it is extraordinarily cheap — $0.10/$0.50, a tenth of Haiku 4.5’s rate. Over 100,000 tokens every rate multiplies by five. It carries a 1M context window, 128K maximum output and adaptive thinking that is on by default; manual budget_tokens now returns a 400. The practical trap is in the next section.

Fable 5.1 is the ceiling at $10/$50 with a $0.25 cache read — worth it only for a cache-dominated agent that keeps a very large context resident, or for work that genuinely needs the top of the range. It is a covered model, with the retention consequences described below.

Sonnet 4.5 retires on 30 November 2026, with Sonnet 5.5 as the named replacement. Nothing changes on a subscription; this is an API lifecycle event. The list price falls from $3/$15 to $2/$10, but Sonnet 5.5 runs the newer tokenizer, so the effective saving on identical text is closer to 13% than 33%. Haiku 4.5 is still Active with a retirement floor of “not sooner than 15 October 2026” — a floor, not a date, and several third-party lifecycle trackers render it wrongly as a hard deadline. Anthropic’s 60-day notice commitment makes a 15 October retirement impossible in principle; no notice has been issued. Dates for every model are in our retirement calendar.

The practical rule: Sonnet 5.5 for most things, Opus 5.5 when the work is hard, Haiku 5.5 for high-volume short-prompt work, Fable 5.1 only for the ceiling.

What the API actually bills you

Four behaviours move real invoices and appear nowhere on the rate card. All four are subscription-irrelevant: Pro, Max and Team are not metered per token.

The tokenizer adds about 30%. Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text; Sonnet 4.6, Haiku 4.5 and earlier use the previous one. Every per-token comparison against an older Claude, or against an OpenAI or Google model, understates Claude’s cost until you normalise for it. Our API cost calculator applies each vendor’s published characters-per-token figure for this reason.

Haiku 5.5’s 100,000-token line is steeper than anyone else’s, and the tokenizer moves it. OpenAI charges 2x input over 272K tokens; Google and xAI charge 2x over 200K. Anthropic charges 5x over 100,000 — the steepest long-prompt surcharge in the market, at the lowest threshold. Worse, the threshold is denominated in the new tokenizer’s tokens, so text that measured 100,000 tokens on Haiku 4.5 measures about 130,000 on Haiku 5.5: the cliff bites at roughly 77,000 Haiku-4.5-equivalent tokens. A cached agent loop sitting on 120,000 tokens of resident context pays $0.006 a turn in cache reads; trim it to 95,000 and the same loop pays $0.00095 — 84% less for 21% less context. That is the cheapest 25,000 tokens you will ever delete, and it is the whole reason to read the Haiku 5.5 pricing analysis before routing volume traffic to it. Every vendor’s threshold, normalised to the same document, is in our long-prompt cost guide — on the same text Anthropic’s cliff arrives four times sooner than OpenAI’s, not 2.7 times sooner as the raw token counts imply.

A refused request can cost money. Since 24 September 2026 Anthropic bills pre-output refusals in three of five categories — bio, frontier_llm and reasoning_extraction — at the full input rate of the model that ran them, returning an empty content array. cyber, general_harms and uncategorised refusals remain free. The detail that catches teams out is that reasoning_extraction is not a harm category: it describes a request asking the model to write its internal reasoning into the response, which means long-standing “show your work” and “explain your reasoning step by step” instructions sitting in system prompts can now generate billed, empty responses. Where fallback is enabled the refused attempt is billed in addition to the retry.

Two usage fields stopped being the bill. When on-demand compaction fires, usage.input_tokens and usage.output_tokens exclude the compaction iteration; the real total lives in a new usage.iterations array. In Anthropic’s own worked example the two obvious fields report 23,000 input tokens against 203,000 billed — roughly 8x. Nothing is concealed and no rate moved, but any cost dashboard or chargeback script reading the two obvious fields undercounts, silently, and only on the requests that cost the most. Sum usage.iterations when it is present.

Two smaller multipliers stack on all of the above: pinning inference to the US with inference_geo: "us" costs 1.1x on Claude 4.6 and later, and regional or multi-region endpoints on Bedrock and Google Cloud carry a 10% premium over global ones. There is no EU equivalent on the first-party API, so European buyers who need residency are choosing between the two marketplaces — Bedrock and Vertex compared, where the feature sets and the model lists differ more than the prices do.

Retention, if you are regulated. Anthropic’s covered models — the Mythos class, including Fable 5 and Fable 5.1 — are retained for 30 days on every platform where they are offered, are not available with zero data retention, and cannot be used under a HIPAA Business Associate Agreement at all. Non-covered models, including Sonnet 5.5 and Opus 5.5, are unaffected and remain available under a BAA. The Life Sciences Verification Program relaxes biology classifiers for vetted organisations, but excludes BAA-enabled orgs, third-party platforms and individual Pro or Max plans, and charges mandatory 30-day retention of flagged traffic — it compounds the BAA caveat rather than resolving it.

What Claude genuinely does better

Writing that sounds like a human wrote it. This is Claude’s most defensible advantage and the reason most professional writers stay subscribed. Run the same brief through ChatGPT and Claude and Claude’s output is noticeably less “AI-voice” — fewer hedges, fewer symmetrical paragraphs, fewer predictable transitions. The difference is subtle per document and large per month, measured in editing avoided.

Long context that holds up. The 1M-token window on current models is included at standard pricing, and quality degrades more slowly as it fills than it does on rivals. Dropping an entire codebase or a 200-page contract in and getting coherent analysis back is the thing Claude users do that users of other assistants mostly do not.

Reasoning you can audit. Given a hard logic problem or a constraint-heavy writing task, Claude works through it visibly before answering, and tends to volunteer the cases where its answer would break. Opus is especially good at this.

Coding with explanations. Claude does not just write code, it explains what it did and why — valuable for review, for learning, and for understanding an unfamiliar codebase. Claude Code is the separate terminal agent and the most-loved AI coding tool in JetBrains’ 2026 developer survey.

One surface instead of several. Anthropic folded Cowork and Design into ordinary conversations in September 2026 and launched Claude Docs and Claude Slides: documents private by default, editable in place, exportable to Google Docs or Word, with presentations downloadable as PowerPoint or PDF. No price changed. The real shift is that agent behaviour — file access, connectors, actions — is now the default behaviour of the chat box rather than a separate product with its own rollout decision. Claude asks before acting by default. Pair that with the 27 integrations (Stripe, Square, Xero, Gusto, Shopify, Salesforce, Zapier among them) that shipped with Claude for Small Business on all paid plans.

Projects. Saved system instructions, reference documents and context, reusable across sessions: one project per writing voice, per client, per codebase. It removes most of the setup tax from recurring work.

Where Claude falls short

No image generation. The single biggest reason serious users keep ChatGPT alongside Claude. Claude analyses images; it cannot create them.

No voice mode. No spoken conversations, no dictation, no hands-free thinking. For anyone who has built a workflow around ChatGPT’s voice feature, this is a real gap.

No video generation. Text and code only.

Weak web research. Claude can search the web, but it is slower and less polished than Gemini or Perplexity. For research that needs current information with reliable citations, use something else.

A smaller ecosystem of third-party add-ons. Artifacts are excellent; ChatGPT’s Custom GPT store has orders of magnitude more community-built tools. If you need “a GPT for X” and X is niche, ChatGPT wins on selection.

Reliability you should not design around. Anthropic logged 21 status incidents between 1 and 24 August 2026, including one critical and six major, and the only service tier that published an uptime target has been withdrawn from sale — it never covered Sonnet 5 or Opus 5 in any case. This is not a reason to avoid Claude; over the same window OpenAI logged more total incident-hours, and busy status pages are the industry norm. It is a reason not to make Claude a single point of failure in anything automated. Keep a second model family qualified against your own evals.

Claude vs ChatGPT vs Gemini

• Writing: Claude > Gemini > ChatGPT. The gap is meaningful, not marginal.
• Coding: Claude > ChatGPT ≈ Gemini. Claude Code has pulled well ahead of Copilot on complex tasks.
• Research with citations: Gemini > Perplexity > ChatGPT > Claude. Claude’s weakest feature.
• Images, video, voice: ChatGPT > Gemini > Claude. Claude is not in the running.
• Everyday quick questions: ChatGPT > Claude. Breadth wins.
• Anything you will publish under your name: Claude.

The full head-to-head is in Claude vs ChatGPT; for assistants generally, see the best AI chatbots.

The Claude Code factor

Claude Code, the terminal coding agent, is a separate product with its own value proposition and its own usage limits. JetBrains’ April 2026 developer survey put it at 46% “most loved” among AI coding tools — 27 points ahead of Cursor and 37 ahead of GitHub Copilot — and its primary-tool share overtook Cursor’s during the year.

If you are an engineer, the right way to read Claude’s pricing is: Pro for chat plus Claude Code at entry limits, Max 20x if the agent is central to how you work. That is where the ladder’s one discount lives, and it is the most common professional AI stack among engineers in 2026. Note that Claude Code’s weekly limits were cut 17% in September 2026 against a base Anthropic has never published; see the Claude Code review for what the limits actually permit.

Procurement notes that are not about the product

None of this changes what Claude does. All of it changes sentences that appear in vendor questionnaires.

Independent evaluation is now version-specific, not vendor-wide. Anthropic named Accenture, through Faculty, as its first embedded evaluator in September 2026, each side expecting to invest at least $1bn over five years, with employee-comparable access — but Anthropic funds the work directly, and the original commitment’s publication right (no redaction for being unfavourable) does not appear in the announcement. Separately, Anthropic withheld Mythos 5.1 from the UK AI Security Institute at the White House’s request. If your risk documentation cites third-party evaluation as a control, that claim is now specific to a model version.

A Claude contract is getting harder to leave. The Claude Marketplace lets eligible customers spend “a portion” of committed Anthropic spend on partner products, with Anthropic consolidating the invoicing. AWS publishes its equivalent cap at 25%; Anthropic has published no percentage, no eligibility threshold and no fee structure. Treat it as unbudgetable until someone puts a number in writing — and note that a marketplace listing is distribution, not assurance: Anthropic’s own directory states it cannot verify what plugins contain.

On governance there is currently nothing to choose between Claude and ChatGPT. Both vendors signed the White House Accord on Super Intelligence on 29 September 2026, under which companies select their own auditors, need not name them, face no deadline and need not publish findings; both are named in the same FTC investigation confirmed a day later. An open investigation is not a finding.

Federal buyers are a separate case. A federal judge vacated Anthropic’s “supply chain risk” designation in August 2026 and permanently enjoined its enforcement, but the practical position remains contested pending appeal, and Claude is still absent from the Department of War’s GenAI.mil portal where ChatGPT Mil, Grok for Government and Gemini are live. Commercial pricing, availability and terms have never been affected.

Who should use Claude

• Professional writers, editors and content creators — the output quality difference is worth paying for
• Software engineers — especially for refactoring, code review and architectural decisions
• Researchers and academics working across long documents
• Consultants, lawyers and analysts doing nuanced written work
• Anyone who values output quality over feature breadth

Who shouldn’t use Claude as their primary AI

• People who need image or video generation — use ChatGPT or a specialist tool
• People who want voice conversations — ChatGPT is the only mainstream option
• Casual users asking one question a day — the free tier suffices, but ChatGPT’s free tier is slightly better
• Anyone depending on real-time web research — Gemini or Perplexity do it better

FAQ

Is Claude Pro worth $20 a month?

For writing or coding work, yes, and it is the better of the two $20 assistant subscriptions. Pro adds Opus access, Projects, the document and slide surfaces and Claude Code at entry limits. If your use is occasional question-answering, the free tier is enough and ChatGPT’s free tier is slightly more generous.

Should I upgrade from Pro to Max?

Only if you are hitting caps. Max 5x costs exactly the same per unit of usage as Pro — five times the price for five times the headroom — so it is a capacity decision, never a value one. Max 20x is different: at $200 for 20x it buys usage at half the rate every other rung charges, which makes it the right tier for anyone running Claude Code all day.

Which Claude model should I call on the API?

Sonnet 5.5 at $2/$10 for most work, Opus 5.5 at $4/$20 when the task is genuinely hard, Haiku 5.5 at $0.10/$0.50 for high-volume short-prompt traffic, Fable 5.1 at $10/$50 only for the ceiling. Price the specific workload in our API cost calculator — the tokenizer difference alone moves cross-vendor comparisons by about 30%.

Why did my Haiku 5.5 bill jump without a rate change?

Almost certainly because your prompt crossed 100,000 tokens, where every Haiku 5.5 rate multiplies by five. The threshold is measured in the newer tokenizer’s tokens, so it arrives sooner than an older token count suggests — around 77,000 Haiku-4.5-equivalent tokens. Trimming resident context back under the line is the single highest-leverage change available on that model — a 21% cut in context takes 84% off a cached turn. Note that prompt caching does not help you here: Anthropic counts cache reads toward the threshold, so a 150,000-token prompt that is almost entirely a cache hit still pays the higher rate on every token. Thresholds for all four vendors are compared in our long-prompt cost guide.

Do I need to migrate off Claude Sonnet 4.5?

Yes, if you call claude-sonnet-4-5-20250929 directly or through a platform that pinned it: it retires on 30 November 2026, with Sonnet 5.5 as the replacement. Note that temperature, top_p and top_k return a 400 on Sonnet 5.5 and raise a TypeError in the Python SDK from v1.0, and that Sonnet 4.5’s extended-thinking mode is not accepted on models after Sonnet 4.6.

Does Claude generate images or video?

No. Neither, on any plan. It analyses images but cannot create them, and it has no video or voice generation at all. This is the main reason many Claude subscribers also keep ChatGPT.

The verdict

Claude is the default AI for serious work in 2026. For writing, for code you will ship, and for thinking through a hard problem, the output quality difference is real, and the $20 Pro plan is among the highest-leverage tool subscriptions in the category.

The broader point is that the assistant market has split. ChatGPT is the default mass-market tool, genuinely good at many things and still where most people should start. But a quiet migration has happened among people who use AI professionally — writers, engineers, researchers, strategists have largely landed on Claude, not because the work is smarter but because it rewards output quality, and that is where Claude pulls ahead.

If you have been paying $20 for ChatGPT Plus and your work is mostly writing or coding, run Claude Pro alongside it for a month. Two weeks of side-by-side use is usually enough to decide, and most people who run the experiment do not go back.

What changed

• 15 Oct 2026: Google priced Gemini 4 Argon at Opus 5.5’s exact $4/$20 from 1 December; Artificial Analysis scores Opus 5.5 at 58 against Argon’s 53, and Sonnet 5.5 at 56. Rate card re-verified unchanged.
• 11 Oct 2026: Corrected the Team seat multipliers against Anthropic’s Team plan documentation — a Standard seat is 1.25x Pro per session and Premium 6.25x, not 1x and 5x, so both cost $20 per 1x rather than $25. Added the monthly API credit every paid plan but Pro returns.
• 8 Oct 2026: Haiku 5.5 launched 7 Oct at $0.10/$0.50 below 100K-token prompts and 5x that above — the steepest long-prompt surcharge in the market; Sonnet 5.5’s cache read was cut to $0.10 the same day. See the analysis.
• 8 Oct 2026: Plan list corrected against Anthropic’s own pricing page — Team’s minimum is 2 seats, not 5, and Enterprise is $20/seat plus API-rate usage, not “custom”.
• 1 Oct 2026: Sonnet 4.5 deprecated; retires 30 Nov 2026 with Sonnet 5.5 as replacement. Haiku 4.5’s “15 October” floor is not a retirement date.
• 1 Oct 2026: Anthropic and OpenAI signed the White House accord and were named in the same FTC probe — no basis to choose between them on governance.
• 28 Sep 2026: Sonnet 5.5 shipped at Sonnet 5’s rates with six migration gotchas; it is now the recommended default.
• 27 Sep 2026: Pre-output refusals are billed again in the bio, frontier_llm and reasoning_extraction categories.