The 2026 lineup
Anthropic shipped four models in two months this summer, and the naming changed with them - Claude 4.x gave way to a named-tier system. The good news is that the lineup is genuinely simpler than OpenAI's nine-model picker. The bad news is that one member of it is now conspicuously stale.
| Model | Released | Context / output | API price | Built for |
|---|---|---|---|---|
| Claude Fable 5Frontier | 9 Jun 2026 | 1M / 128k | $10 / $50 | Hardest reasoning and long-horizon agent runs. Slowest and most expensive. |
| Claude Opus 5Recommended default | 24 Jul 2026 Newest | 1M / 128k | $5 / $25 | Complex agentic coding and enterprise work - near-Fable quality at half the price. |
| Claude Sonnet 5Workhorse | 30 Jun 2026 | 1M / 128k | $2 / $10 | The speed-to-intelligence sweet spot; what most production traffic should use. |
| Claude Haiku 4.5Fast tier | Oct 2025 | 200k / 64k | $1 / $5 | Cheapest and fastest, but a Feb 2025 knowledge cutoff and no 1M context. Ageing |
| Claude Mythos 5Restricted release | 9 Jun 2026 | 1M / 128k | $10 / $50 | Identical to Fable 5 but ships without safety classifiers; available only through Project Glasswing. |
Fable 5, Opus 5 and Sonnet 5 all run adaptive thinking that cannot be switched off, with effort control and task budgets to bound it. All three support code execution, vision, tool use, memory tools, and context editing and compaction - the last of which is what makes a 1M-token window practical rather than theoretical. A Batch API beta pushes output up to 300k tokens.
Opus 5 is the one to default to. Anthropic's own guidance now points there rather than at Fable 5, and the pricing gap is the reason: on CursorBench 3.2 at maximum effort Opus 5 lands within 0.5% of Fable 5's peak score at half the cost per task, and on OSWorld 2.0 it beats Fable 5 outright at roughly a third of the cost.
The benchmark picture
Claude Opus 5 figures. ARC-AGI 3 is the standout - roughly three times the next-best score - though 30% also shows how far novel-puzzle reasoning still has to go.
The honest counterweight: on the Epoch Capabilities Index, a composite measure, Opus 5 trails both Fable 5 and GPT-5.6. On Apex SWE, Fable 5 leads by a small margin. On the GDP-weighted Vales Index the entire top tier sits within about 1% of each other. Anyone telling you one model is decisively ahead in August 2026 is selling something.
Claude Code, Cowork and the agent surfaces
Anthropic's product strategy is unusual: rather than one app that does everything, it ships separate surfaces for separate kinds of work.
- Claude Code - the terminal agent that made Anthropic's reputation with developers. Since 6 August 2026 Teams and Enterprise can run it on self-hosted infrastructure in public beta, which removes the main blocker for regulated engineering orgs.
- Claude Cowork - generally available since 9 April 2026, this is Claude Code's logic pointed at ordinary office work: folder-scoped file access, persistent project instructions, scheduled recurring tasks, and - crucially - the ability to steer a task mid-run rather than restarting it. It reached web and mobile on 7 July (Max first) and gained a Chrome side panel with cross-device session continuity on 12 August. Anthropic's own usage data shows most Cowork users are not coding.
- Claude Tag - the Slack integration, upgraded on 13 August to be roughly 30% better at judging when it should actually respond, which is the whole ballgame for a bot in a busy channel.
- MCP - the 2026-07-28 revision moved to a stateless architecture with OAuth/OIDC hardening plus Apps and Tasks extensions. MCP is now the de facto standard for tool connectivity across the industry, and it is Anthropic's.
On the enterprise side the last quarter has been dense: self-serve HIPAA configuration (14 July), FedRAMP High authorisation for Government Desktop (7 July), inference hooks for real-time DLP across chat, Code and Cowork (5 August), skill and plugin security scanning (6 August), and a Compliance API extended to Cowork and Code sessions (11 August). That is a company selling to CISOs, and it shows.
Plans, prices and the limits problem
| Plan | Price | What you get | Best for |
|---|---|---|---|
| Free | $0 | Chat, web search, memory, code execution, file creation, extended thinking and MCP connectors at 200k context, with daily limits | Trying it properly |
| Pro | $20$17 annual | Adds Claude Code, Cowork, unlimited projects, research mode and the Microsoft 365 connector | Most people Sweet spot |
| Max 5× | $100 | 5× usage, higher output limits, priority access, early features | Daily Claude Code users |
| Max 20× | $200 | 20× usage - the only tier where long agent runs stop being rationed | Heavy agentic workloads |
| Team | $20–25per seat, min 5 | SSO and domain capture, central billing, admin controls, no training on your data by default. Team Premium at $100–125/seat adds a 5× multiplier | Small teams |
| Enterprise | $20/seat+ usage | SCIM, audit logs, HIPAA-ready, Compliance API, custom retention, EU data residency | Regulated orgs |
The three traps worth knowing before you subscribe. First, Pro and Max carry weekly caps on top of per-session five-hour windows - burn through by Wednesday and you wait. Second, Claude Code draws from the same pool as chat, so a morning of conversation shortens your afternoon of coding. Third, if you have ANTHROPIC_API_KEY set in your environment, Claude Code bills at API rates even with an active subscription - a genuinely expensive surprise. On Team plans, usage credits for members who exceed their seat allowance bill at API rates and are not capped by default.
One more piece of quota arithmetic: Opus consumes your allowance roughly 1.7× faster than Sonnet. If you are hitting walls on Max, switching routine work to Sonnet 5 is usually a bigger win than upgrading.
Where Claude is strongest
- Prose. This is the least contested claim in the category. Claude's default register - fewer hedges, less listicle instinct, better paragraph rhythm - needs less editing than anything else on the market. If a human reads your output end-to-end, start here.
- Agentic coding. ~97% on SWE-bench Verified, and Claude Code remains the reference implementation of a terminal coding agent. Self-hosted deployment closes the last enterprise gap.
- 1M tokens as standard. Not a premium add-on - Fable 5, Opus 5 and Sonnet 5 all carry it, with context compaction that makes it usable in practice.
- Steerable long-running work. Cowork lets you redirect a task mid-execution. Sounds minor; it is the difference between a 40-minute agent run being useful and being wasted.
- Compliance depth. HIPAA, FedRAMP High, SCIM, audit logs, DLP hooks, plugin security scanning, EU residency on Enterprise.
- Price-performance at the top. Opus 5 at $5/$25 undercuts GPT-5.6 Sol's $5/$30 while matching it on most published benchmarks.
Where it falls short
- The usage meter is the worst in the category. Weekly caps plus session caps plus a shared chat/code pool, and limit transitions that can drop you into pay-per-token without clear warning.
- No image or video generation. None. If your workflow needs a diagram rendered or a picture made, you are opening another tab.
- Haiku 4.5 is the weak link. A February 2025 knowledge cutoff and 200k context in a lineup where everything else has 1M looks increasingly odd.
- Thinner integration catalogue. Documented gaps include Salesforce, Shopify, Square, Toast and NetSuite - awkward for exactly the small businesses Anthropic has started courting.
- Refusals need engineering around. Fable 5 can return
stop_reason: "refusal"as a normal 200 response. You are not billed, but production systems need fallback logic - and false-positive refusals remain a real, if reduced, annoyance. - Still hallucinates in agent mode. Third-party testing of Cowork found fabricated statistics and misattributed citations. Long autonomous runs need checking, not trusting.
- EU residency is Enterprise-only. Everything below that defaults to US infrastructure.
- Consumer mindshare. ChatGPT has roughly 900M weekly users; Claude is a fraction of that, which means fewer third-party integrations get built for it first.
Claude vs the alternatives
The top of the market is close enough now that fit matters more than leaderboard position. Here is the field as it stands in August 2026:
| Tool | Flagship model | Price | Where it wins | Rating |
|---|---|---|---|---|
| ClaudeAnthropic | Claude Fable 5 · Opus 5Opus 5: 24 Jul 2026 | Free / $20 Pro / $100–200 Max | Writing and code. The cleanest default prose in the category, ~97% SWE-bench Verified, 1M context across the lineup, and steerable long-running agents. | 4.5/5 |
| ChatGPTOpenAI | GPT-5.6 Sol9 Jul 2026 | Free / $8 / $20 / $100–200 | Breadth and ecosystem. Image, video, voice, custom GPTs and the deepest integration catalogue, plus the strongest long-horizon web research (BrowseComp 90.4%). | 4.6/5 |
| GeminiGoogle | Gemini 3.1 Pro19 Feb 2026 | Free / $4.99 / $19.99 / Ultra | Distribution and price. 1M context, 77.1% on ARC-AGI-2, already inside Gmail, Docs, Android and NotebookLM - and the cheapest credible paid tier anywhere. | 4.5/5 |
| GrokxAI | Grok 4.3updated 1 Aug 2026 | Bundled with X Premium+ | Token price. $1.25 / $2.50 per million undercuts every frontier rival, with live X data and a looser content policy. | 4.0/5 |
| Microsoft CopilotMicrosoft | Multi-modelincl. DeepSeek V4 trials | Bundled with Microsoft 365 | Proximity to the work. Native in Word, Excel, Teams and Windows with tenant governance IT already runs. | 4.0/5 |
| DeepSeekDeepSeek | DeepSeek V4 | Free app; commodity API | Cost and control. Open weights you can self-host at a fraction of frontier pricing - good enough that Microsoft has trialled it inside Copilot. | 4.1/5 |
| PerplexityPerplexity AI | Multi-model + Comet | Free / $20 Pro / Max | Sourced answers. Citation-first research where every claim is checkable - the cleanest surface for work you have to defend. | 4.2/5 |
| Le ChatMistral AI | Mistral open-weight family | Free / low-cost Pro | European data residency. Open weights and EU hosting without a US cloud dependency - a procurement answer more than a capability one. | 3.8/5 |
What the table actually tells you
Claude and ChatGPT have swapped strengths since 2024. ChatGPT used to write better and Claude used to reason better; today it is closer to the reverse. Claude's edge is the quality of a single artefact - a document, a pull request, a contract analysis. ChatGPT's edge is the number of different things it can do without you leaving the tab.
The half-star gap to ChatGPT is about ecosystem, not intelligence. On published benchmarks the two are within a percentage point or two almost everywhere. ChatGPT scores higher here because it also generates images, makes video, runs voice, has a marketplace, and integrates with more of the software people actually use.
Gemini is the value play and the context rival. Google AI Plus at $4.99 with 1M context is hard to argue with on price, and for anyone living in Gmail and Docs the integration can outweigh a benchmark gap entirely.
Which one should you actually pay for?
Claude Pro - $20
The least editing required per thousand words. This is the category Claude wins outright.
Claude Max 5× - $100
Claude Code is the reference terminal agent, but $20 Pro will not survive daily use. Budget for Max.
ChatGPT Plus
Images, video, voice, custom GPTs and the widest integration catalogue for the same $20.
Gemini / Google AI Plus
$4.99 for 1M context and native Workspace integration. Nothing else is close on price.
Claude Sonnet 5
1M context at $2/$10 makes whole-corpus analysis genuinely affordable at volume.
Claude Enterprise
HIPAA, FedRAMP High, SCIM, DLP hooks, EU residency and a Compliance API covering agent sessions.
Perplexity Pro
Citation-first by design, so you verify rather than trust.
Grok 4.3 or DeepSeek V4
An order of magnitude cheaper per token, and for extraction or classification the gap barely registers.
Pros and cons
What we liked
- Best default prose quality of any frontier model
- ~97% SWE-bench Verified; Claude Code is the reference coding agent
- 1M-token context standard across Fable 5, Opus 5 and Sonnet 5
- Opus 5 at $5/$25 - near-frontier quality at half Fable 5's price
- ARC-AGI 3 at 30%, roughly triple the next-best model
- Cowork tasks can be steered mid-run instead of restarted
- Deep compliance stack: HIPAA, FedRAMP High, SCIM, DLP, Compliance API
- MCP is now the industry standard for tool connectivity
What frustrated us
- Weekly caps stacked on five-hour session caps, with opaque transitions
- Claude Code and chat share one quota pool
ANTHROPIC_API_KEYsilently overrides your subscription and bills API rates- Team usage credits bill at API rates, uncapped by default
- No image or video generation at all
- Haiku 4.5 stuck at 200k context and a Feb 2025 cutoff
- Integration gaps: Salesforce, Shopify, Square, Toast, NetSuite
- Cowork can still fabricate statistics and citations on long runs
- EU data residency requires Enterprise
FAQ
What is the newest Claude model?
Claude Opus 5, released 24 July 2026. It is now Anthropic's recommended default - $5/$25 per million tokens, a 1M-token context window and 128k max output, landing close to the more expensive Fable 5 on most benchmarks and beating it on OSWorld 2.0 at about a third of the cost.
Should I use Fable 5 or Opus 5?
Opus 5 for almost everything. Fable 5 is worth the doubled price only for the hardest long-horizon reasoning, and it retains an edge on cybersecurity exploitation tasks. On the composite Epoch Capabilities Index, Fable 5 still leads Opus 5 - but not by enough to justify $10/$50 for routine work.
Is Claude Pro worth $20 a month?
For writing and light coding, yes - Pro unlocks Claude Code, Cowork, unlimited projects and research mode. If you use Claude Code daily you will hit the weekly caps, and Max 5× at $100 becomes the honest price of that workflow.
Claude or ChatGPT?
Claude for writing, code and long documents; ChatGPT for breadth. ChatGPT also generates images and video, runs a plugin marketplace and integrates with more third-party software. Many people pay for both - at $40 a month combined that is still cheaper than one Max plan.
Can Claude generate images?
No. Claude reads and analyses images but does not create them, and there is no video generation either. For visual output you need ChatGPT, Gemini or a dedicated tool.
What is Claude Cowork?
An agent surface for non-coding work, generally available since 9 April 2026 and now on desktop, web, mobile and a Chrome side panel. It gets folder-scoped file access, keeps persistent project instructions, can run scheduled recurring tasks, and can be redirected mid-task. It is included on all paid plans at no extra cost.
Does Anthropic train on my data?
Not on Team and Enterprise plans - that is the default. Enterprise adds SCIM, audit logs, custom retention, a Compliance API covering Cowork and Claude Code sessions, and EU data residency, which is not available on lower tiers.
Why did my Claude Code bill spike?
Almost certainly an ANTHROPIC_API_KEY set in your shell environment. Claude Code will use it and bill at API rates even while you hold an active Pro or Max subscription. Unset it, or check your organisation's usage-credit settings if you are on Team.
Verdict
Claude is the model to reach for when the output matters more than the variety of outputs. It writes better than anything else without prompting gymnastics, it is at the frontier on agentic coding, and it gives every current model a million tokens of context rather than reserving that for a premium tier. Opus 5 in particular has changed the value calculation - near-Fable performance at half the price is the most consequential thing Anthropic shipped this year.
The reasons it lands at 4.5 rather than higher are all commercial rather than technical. The usage meter is the most confusing in the category and occasionally expensive by surprise. There is no image generation, no video, and a visibly thinner integration catalogue. Haiku 4.5 needs replacing. And ChatGPT's ecosystem advantage compounds in ways benchmark parity does not fix.
If you write or code for a living, Claude Pro is the best $20 in AI - and step up to Max before you get annoyed rather than after. If you need one tool that does everything, this is not it, and that is a deliberate choice on Anthropic's part rather than a failure.
The specialist's choice, in a market that rewards generalists
Best-in-class writing, frontier coding, 1M context as standard and the deepest compliance stack going. Held back by an opaque usage meter, no image or video generation, and an integration catalogue that trails OpenAI's.