Models
Which model to run Claude Code on — what each costs, what each is for, what changed with Opus 5, and how the open-weight tier now compares.
The current lineup
Which Claude model to reach for, and what each one costs.
| Model | Model ID | Released | Context | Price (in / out) | Best For |
|---|---|---|---|---|---|
| Claude Fable 5 | claude-fable-5 | Jun 2026 | 1M tokens | $10 / $50 per 1M | Hardest reasoning and long-horizon agentic work |
| Claude Opus 5 | claude-opus-5 | Jul 2026 | 1M tokens | $5 / $25 per 1M | Agentic coding, deep reasoning, long autonomous runs |
| Claude Opus 4.8 | claude-opus-4-8 | May 2026 | 1M tokens | $5 / $25 per 1M | Previous Opus generation — pin it if you need stable behaviour |
| Claude Sonnet 5 | claude-sonnet-5 | Jul 2026 | 1M tokens | $3 / $15 per 1M | High-volume coding and agentic work at lower cost |
| Claude Haiku 4.5 | claude-haiku-4-5 | Oct 2025 | 200K tokens | $1 / $5 per 1M | Fast, simple, latency-sensitive tasks |
Sonnet 5 carries an introductory rate of $2 / $10 per 1M tokens through 31 August 2026. Model IDs are the exact strings the API expects — they take no date suffix.
What's new in Opus 5
Released 24 July 2026 at the same price as Opus 4.8 — $5 in, $25 out per million tokens.
Anthropic describes Opus 5 as a thoughtful, proactive model that is much stronger at verifying its own work and iterating carefully — the behaviour that matters most for long agentic runs. It is also billed as their most aligned model to date. Two things carry over into how you drive it from Claude Code: thinking is now on by default (on Opus 4.8, omitting the thinking parameter meant no thinking), and disabling thinking is only permitted at effort high or below.
| Evaluation | Reported result |
|---|---|
| Frontier-Bench v0.1 | Surpasses all other models; more than doubles Opus 4.8 |
| ARC-AGI 3 | Three times the score of the next-best model |
| CursorBench 3.2 | Within 0.5% of Fable 5's peak, at half the cost per task |
| OSWorld 2.0 | Beats Fable 5's best result at just over a third of the cost |
| Zapier AutomationBench | Pass rate around 1.5× the next-best model |
| Organic chemistry (internal) | 10.2 percentage points above Opus 4.8 |
| Protein tasks (internal) | 7.7 percentage points above Opus 4.8 |
Anthropic published these as charts and comparisons rather than raw scores, so no absolute percentages are available yet. Where it still trails: cyber-exploitation work (behind Mythos 5) and long-running autonomous biology research. Source: Anthropic — Introducing Claude Opus 5.
Claude Fable 5
Anthropic's most capable widely released model — back online since 1 July 2026, on plan-gated access.
Status: available. Included as standard on Max, on Team premium seats and on premium seats of seat-based Enterprise plans — up to 50% of your weekly usage limits at no extra cost, then pay-as-you-go usage credits. Pro and standard Team seats reach it on usage credits from the first token; standard Enterprise seats only if the organisation enables credits, and usage-based Enterprise plans bill at standard API rates. This structure took effect 20 July 2026.
Why it was gone: Anthropic pulled Fable 5 on 12 June 2026, three days after launch, when a US export-control directive required restricting access by nationality — something it had no reliable way to verify in real time. The controls lifted on 30 June and the model was redeployed on 1 July.
Fable 5 sits above the Opus tier for the most demanding reasoning and long-horizon agentic work, at $10 in / $50 out per million tokens with a 1M-token context window. Claude Mythos 5 is the same model family with identical pricing and API behaviour, available only through Project Glasswing.
What makes it different to drive
- Thinking is always on. There is no thinking parameter to configure — sending one is rejected. Depth is controlled entirely through the effort setting.
- The raw chain of thought is never returned. You get a readable summary or nothing at all, depending on the display setting.
- Requests can be declined outright. Safety classifiers may refuse a request and return a refusal rather than an answer — most often on cybersecurity and research-biology topics. Benign adjacent work occasionally trips them, which is why production callers configure a fallback model.
- 30-day data retention is mandatory. It is not available to organisations configured for zero data retention.
- Turns run long. A single request on a hard task can run many minutes — plan timeouts and progress indicators accordingly.
When a Fable 5 request is declined, the documented fallback target is Claude Opus 4.8 — which is why the Opus line remains load-bearing even for teams standardising on Fable.
The rest of the field
What the other labs shipped in July 2026, and where each one earns its place.
| Model | Price (in / out) | Published results | Best For |
|---|---|---|---|
| GPT-5.6 SolOpenAI | $5 / $30 per 1M | Terminal-Bench 2.1 88.8% · SWE-Bench Pro 64.6% · ExploitBench 1 73.5% · BrowseComp 92.2% | Hardest reasoning, security work, deep agentic runs |
| GPT-5.6 TerraOpenAI | $2.50 / $15 per 1M | Terminal-Bench 2.1 87.4% · Coding Agent Index 77.4 | Everyday production work — the default tier |
| GPT-5.6 LunaOpenAI | $1 / $6 per 1M | Terminal-Bench 2.1 84.7% · MRCR long-context 41.3% | High-frequency, short-context tasks where cost dominates |
| Grok 4.5xAI | X Premium+ / API | Coding Agent Index 76 (Grok Build harness) · beats Opus 4.8 on DeepSWE 1.0 + Terminal-Bench 2.1 | Coding agents where token spend matters |
| Gemini 3.6 FlashGoogle DeepMind | $1.50 / $7.50 per 1M | AA Intelligence Index 50 · 17% fewer output tokens than 3.5 Flash | High-volume agentic coding at Flash-tier cost |
Two things worth knowing.OpenAI's new naming splits generation from tier — the number (5.6) moves with the generation, while Sol, Terra and Luna are durable capability tiers that can advance on their own cadence. Sol is the only tier that unlocks max reasoning effort and ultra mode. And on security specifically, Sol's ExploitBench 1 result of 73.5% is the strongest published security score in this group — Anthropic does not publish an equivalent, and positions Opus 5 as deliberately behind Mythos 5 on cyber-exploitation.
Grok now sits under SpaceX. SpaceX acquired xAI on 2 February 2026 in an all-stock deal valuing the combined business at $1.25 trillion — folding xAI, Grok and the X platform into SpaceX, and now branded SpaceXAI. The merged company listed on the Nasdaq on 12 June 2026. Grok releases and API access run through that entity; the model line keeps the Grok name.
Grok 4.5 ships inside Grok Build, xAI's agentic coding surface, and in Cursor on all plans. It was trained alongside Cursor on real developer session data — debugging traces, multi-file diffs and user corrections — which shows up as token efficiency: xAI reports it resolving SWE-Bench Pro tasks in an average 15,954 output tokens against 67,020 for Opus 4.8 at max effort, a 4.2× gap.
Grok Build is now open source. It is a Rust-based, full-screen terminal agent — codebase understanding, file edits, shell execution, web search and long-running task management — released under Apache 2.0 at github.com/xai-org/grok-build, where it has passed 22k stars. Beyond the TUI it runs headless for scripting and CI, and embeds into editors through the Agent Client Protocol — so the harness is usable independently of which model you point it at.
Open-weight and Chinese-lab models
The open tier closed a lot of ground in 2026 — here is where it actually stands.
| Model | Size / License | Published results | Best For |
|---|---|---|---|
| Kimi K3Moonshot AI · Jul 2026 | ~2.8TOpen weights | GPQA Diamond 93.5% · Terminal-Bench 2.1 88.3% · BrowseComp 91.2% · MCP Atlas 84.2% · HLE (with tools) 56.0% · Frontend Code Arena No.1 (1,679) | Frontend and agentic coding — the strongest open-weight option |
| GLM-5.2Z.ai · Jul 2026 | 753BMIT | Terminal-Bench 2.1 81.0 · SWE-bench Pro 62.1 | Long-horizon autonomous coding on a budget |
| DeepSeek V4-ProDeepSeek · Apr 2026 | 1.6T MoE (~49B active)MIT | SWE-bench Verified 80.6% — highest open-weights, tied with Gemini 3.1 Pro | Cheapest frontier-class coding; self-hosting without vendor contact |
Running Kimi K3 for coding
K3 is the strongest open-weight coding model published to date, but at roughly 2.8 trillion parameters it is not a model most teams will self-host. Moonshot positions it behind Claude Fable 5 and GPT-5.6 Sol on overall performance while ahead of every other model in its own evaluation suite — including Claude Opus 4.8 and GPT-5.5 — across coding and agentic benchmarks. Three practical routes:
- Hosted API. Fastest path, and how most people will use it. Available through Moonshot directly and via aggregators such as OpenRouter, which is the usual way to point an existing OpenAI-compatible client at it.
- Aggregator routing. Any tool that speaks an OpenAI-compatible endpoint — including most terminal coding agents — can target K3 by swapping the base URL and model ID, with no code change.
- Self-hosting. Realistic for DeepSeek V4-Pro and GLM-5.2 (both MIT, weights on Hugging Face, commercially usable without contacting the vendor). For K3, the parameter count puts it out of reach of single-node setups.
One caveat that applies to all three. Open weights and the hosted API are different propositions. Downloading MIT-licensed weights and running them on your own compute carries no data-residency question; sending your codebase to a vendor-hosted endpoint in another jurisdiction does. Decide which one you are actually doing before routing production traffic.