As of October 2026, Claude Opus 5.5 is the best Claude model for coding an app. Anthropic says to start with it, and it ranks first on LMArena's WebDev leaderboard at max effort. For small, clear steps, Sonnet 5.5 costs half as much per token. Use Fable 5.1 only when Opus 5.5 at higher effort still falls short.
I'm Will Chen, and I build Dyad, a free, open-source AI app builder that runs on your computer and works with Claude. Most advice on this question assumes you're typing code in Claude Code or an IDE. This post is for building an app by chatting, where the tool picks the effort level (how long the model thinks before it answers) unless you change it. That one setting moves your cost and wait time as much as the choice of model does.
The current Claude models
Anthropic sells four current Claude models. Here they are as of October 19, 2026, with the API list prices from Anthropic's pricing page.
| Model | What Anthropic says it's for | Price per million tokens (input / output) | In Dyad's model picker |
|---|---|---|---|
| Claude Fable 5.1 | Demanding reasoning and long-horizon agentic work | $10 / $50 | Yes |
| Claude Opus 5.5 | Long-running agentic coding and knowledge work | $4 / $20 | Yes |
| Claude Sonnet 5.5 | "The best combination of speed and intelligence" | $2 / $10 | Yes |
| Claude Haiku 5.5 | High-volume, latency-sensitive tasks | $0.10 / $0.50 for prompts up to 100,000 tokens | No |
All four read up to 1M tokens at once (the context window) and write up to 128K tokens per answer, per Anthropic's models overview. Haiku 5.5 pairs well as a sub-agent, one of the small helper tasks a lead model hands off, and Anthropic says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding, so the rest of this post sticks to the other three. You may also see Claude Mythos 5.1. It's the same model as Fable 5.1 with different safeguards, offered only to verified organizations.
If another guide recommends Sonnet 5, Opus 5 or Haiku 4.5, it's a generation behind. Anthropic still sells those, but they aren't current. And if you're wondering whether Claude Sonnet 3.7 or Claude 4 is better for coding, neither is an option anymore. Sonnet 3.7, Sonnet 4 and Opus 4 are retired, and requests to them fail.
Opus 5.5 or Sonnet 5.5: which Claude model is best for coding?
Opus 5.5 wins: clearly at medium effort, and by a small margin at max.
Dyad starts both models at medium, so that's the fair place to compare them. Artificial Analysis runs a general intelligence index (it covers more than coding). At medium effort it scores Opus 5.5 at 51 and Sonnet 5.5 at 41, at $1.34 against $0.48 per task. Sonnet also starts answering far sooner: about 1 second before it starts writing, against about 23 seconds for Opus. So at Dyad's default, Sonnet gives up 10 points for about a third of the cost and a near-instant start.
Turn the effort up and the two meet. At max effort, Sonnet 5.5 scores 56 at $5.46 per task, and Opus 5.5 scores 58 at $5.98. Anthropic's Sonnet 5.5 announcement says Sonnet pairs best with Opus "at lower effort settings, where it costs less per task," and that "at higher settings, it can perform comparably at a similar cost." So if you catch yourself pushing Sonnet to high or max to get the result you want, switch to Opus instead.
For an outside check, LMArena's WebDev leaderboard ranks models by people's votes on web app tasks. In its October 18, 2026 snapshot, Opus 5.5 at max effort is first with 1806 points, and Sonnet 5.5 at xhigh effort is third with 1769. It's the same leaderboard our guide to picking a model in Dyad points to. Anthropic itself says benchmark margins near the top "have become a less reliable guide to real-world differences," so read a 37-point gap as a lean, not a ruling.
Which Claude model to use for each step of building an app
Here's how I'd pick, step by step. It's Anthropic's own task guidance, mapped onto building an app.
| Step | Where I'd start | Why |
|---|---|---|
| Planning a new app, or turning a vague idea into one | Opus 5.5, high effort | Anthropic says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment." Dyad asks a few planning questions and shows you a blueprint before it builds, and that plan shapes everything after it. |
| Adding a well-scoped feature, like a contact form or a settings page | Sonnet 5.5, medium effort | Anthropic calls Sonnet 5.5 "strongest at well-scoped everyday tasks." |
| Fixing an error the preview shows | Sonnet 5.5, medium effort | Fixing bugs is next on Anthropic's list of Sonnet's strengths. |
| Styling the UI | Sonnet 5.5, medium effort | Anthropic credits it with "a sharp eye for design." |
| A big refactor, or a bug Sonnet couldn't fix | Opus 5.5, high effort | Anthropic's pick for complex agentic coding and large refactors. |
Fable 5.1 isn't in the table on purpose. Anthropic says to move to it when Opus 5.5 at xhigh or max effort still falls short on demanding reasoning or long-horizon agentic work. It costs 2.5 times as much as Opus 5.5 per token, and on the WebDev leaderboard it sits fifth at max effort, below both Opus 5.5 and Sonnet 5.5.
Effort matters as much as the model
Anthropic's guide to choosing a model puts it plainly: "Tuning effort is often a better lever than switching models." Dyad offers five levels for each Claude model: low, medium, high, xhigh and max.
On the same Artificial Analysis index, raising Opus 5.5 from medium to high takes it from 51 to 54 for $1.82 per task, with about 40 seconds before it starts answering. Max gets it to 58 for $5.98, and you wait over nine minutes before it starts. That's 4 more points for over three times the price of high, plus a wait you'd sit through on every edit. It's why the table above stops at high.
As of October 19, 2026, Dyad starts all three Claude models at medium. That matches Anthropic's API default for Opus 5.5, but Anthropic's API runs Sonnet 5.5 and Fable 5.1 at high. Either way, the max-effort scores on leaderboards aren't what you get out of the box.
You change effort in the model picker. Since Dyad 1.11.0, effort is set per model, and Dyad remembers your choice for each model and shows the current level next to it.
What a coding session costs on each model
Anthropic's pricing page works through an example coding session with 50,000 input tokens and 15,000 output tokens. We priced that same session at API list prices with no caching:
- Fable 5.1: $1.25
- Opus 5.5: $0.50
- Sonnet 5.5: $0.25
- Haiku 5.5: about a cent
That's arithmetic on Anthropic's example token counts, not a measured Dyad session. Your sessions will be bigger or smaller. The price ratios between the models stay the same.
Cached input matters more than these four lines suggest. In its Opus 5.5 announcement, Anthropic says cache reads (input the model has already seen, billed at a discount) "make up the majority of agentic and coding work costs." They're $0.20 per million tokens on Opus 5.5 and $0.10 on Sonnet 5.5.
If you're comparing with an older model, watch the token count too. Claude 4.7 and later produce about 30% more tokens for the same text, so Sonnet 4.6 at $3 / $15 costs closer to Sonnet 5.5's $2 / $10 than the listed price suggests. For what a whole app costs on your own key, see the bring-your-own-key cost breakdown.
How to use Claude in Dyad
There are four ways to get Claude into Dyad. If you're still choosing between Claude, GPT, Gemini and local models, or want to know how Auto picks across providers, start with the best AI model for vibe coding.
Your own Anthropic API key
Add an Anthropic key in Dyad's settings and you pay Anthropic directly, at the list prices above. Without Dyad Pro, your prompts go straight from your computer to Anthropic, and your bill shows exactly what each model costs. If Anthropic is your only key, Auto runs Opus 5.5. The steps are in setting up AI models. Since 1.19.0, Dyad Pro users can bring their own keys as well.
Your Claude plan, through your own Claude Code
Since Dyad 1.17.0, Dyad can use your Claude subscription. This is experimental. Dyad runs the official Claude Code that you installed yourself (version 2.1.259 or later). You sign in through Anthropic with claude auth login in your terminal, and Dyad never collects your credentials. Then you turn on Claude Code usage in Dyad's model selector.
Claude Code comes with Claude's paid plans, not the Free plan. Usage counts against your plan's limits: a rolling five-hour window plus weekly limits, shared with your Claude chats. The model and effort level you pick change how fast you use them up. As of Anthropic's October 7, 2026 update to its support article on the Agent SDK and third-party apps, "third-party apps" can still use your subscription limits.
If you're on Dyad Pro, this route goes further: we put it at up to 5x the usage you'd get from Dyad Pro credits alone. Agent mode adds up to 1.5 credits per million tokens, and Build, Ask and Plan modes don't use Dyad credits on this route.
OpenRouter
One OpenRouter key reaches Claude and many other models. Dyad's built-in OpenRouter list doesn't include Claude, so add it as a custom model. Use OpenRouter's ID, which has dots (anthropic/claude-opus-5.5), not Anthropic's hyphens (claude-opus-5-5). OpenRouter also runs Opus 5.5 at high effort by default, where Anthropic uses medium, so expect a higher cost per task than the medium numbers above.
Dyad Pro credits
With Dyad Pro ($20 a month), pick Opus 5.5 or Sonnet 5.5 in the model picker and it bills against your credits. The picker marks Fable 5.1 and Opus 5.5 with six dollar signs and Sonnet 5.5 with five. If you stay on Auto, Dyad Pro tries OpenAI, then Anthropic, then Google, and when it lands on Anthropic it uses Opus 5.5.
Whichever route you use, trying the cheaper model first costs you little. Every AI edit in Dyad is a Git commit, so if Sonnet 5.5 makes a mess, restore the previous version and ask again in a new chat on Opus 5.5. Each chat keeps the model it started with, so changing the model in one chat won't change another.
FAQ
What is the best Claude plan for coding?
Claude Pro, at $20 a month (or $17 a month billed annually), is the cheapest Claude plan that includes Claude Code. Max starts at $100 a month and gives you 5x or 20x Pro's usage. On Pro, Fable 5.1 is only available through usage credits. Max includes it up to half of your weekly limits, with a separate weekly Fable limit. If you mostly build on Opus 5.5 and Sonnet 5.5, start on Pro and move up if you keep hitting the five-hour limit.
Is there a free Claude model for coding?
Not on Anthropic's API, which has no free tier. Claude's Free plan lets you chat with Sonnet and Haiku, but it doesn't include Claude Code, so it won't connect to Dyad. Dyad itself is free, so with no Claude budget you can start on a free model through OpenRouter or a local model, then bring Claude in for the hard steps.
Which Claude Sonnet model is best for coding?
Claude Sonnet 5.5, released September 28, 2026. Sonnet 5 costs the same $2 / $10 per million tokens, so there's no reason to pick the older one. Sonnet 4.5 is deprecated and retires on November 30, 2026, after which requests to it fail.
Where to start
Open Dyad, connect Claude whichever way suits you, and build the first version of your app on Opus 5.5. When the next step is small and clear, try it on Sonnet 5.5 at medium. If you can't tell the difference, you've found your everyday model.
Which Claude model are you building with right now, and where has it surprised you?
