Claude Opus 5.5 and GPT-6 Sol have arrived!

Hey everyone, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna today, and all three are already live in Draftbit. Opus 5.5 is the new default for Claude Code, and GPT-6 Sol is the new default for Codex.

TL;DR: Claude Code now starts on Opus 5.5 and Codex now starts on GPT-6 Sol, both more capable and cheaper to run than the models they replace. GPT-6 Luna is a new low-cost option. Also new: browser controls in previews, better project import, and fixes for custom domains, Supabase publishing, and Claude Code.

Claude Opus 5.5: faster, more capable, and cheaper

Opus 5.5 performs at the level of Claude Fable 5.1 on most work, costs 40% less to run than Opus 5, and generates output more than 30% faster. In coding tests, it solved more tasks than Opus 5 while making about 40% fewer tool calls and using half the tokens, and when several Claude models each built a game from a single prompt, Opus 5.5 scored highest for graphics and polish.

On Draftbit credits, input and output tokens cost 20% fewer credits than on Opus 5, and cache reads (the context the agent re-reads on every step, which makes up most of the cost of coding work) cost 60% fewer. How many tokens a task takes still depends on the task, so your savings will vary.

GPT-6 Sol and GPT-6 Luna

GPT-6 Sol is OpenAI's new model for complex coding and agent work, trained with similar methods to GPT-6 Astra. On AutomationBench, a test of business workflows across apps, GPT-6 Sol at extra-high effort beat Claude Opus 5 at max effort for 9% of the cost per task. It also makes about half as many mistakes as GPT-5.6 Sol on OpenAI's factuality evaluation, and OpenAI cut prices for both new models to half of GPT-5.6's promotional pricing. On Draftbit credits, GPT-6 Sol costs the same per input token as GPT-5.6 Terra, the previous Codex default, and about 17% fewer credits per output token.

GPT-6 Luna is the low-cost option, and it's no lightweight: at max effort it beats GPT-5.6 Sol at medium effort for a tenth of the cost. Each token uses a twentieth of GPT-6 Sol's credits, which makes it a good fit for quick edits and routine work. Pick it from the model menu when you start a thread.

What changes for you

When you start a new thread, Claude Code now uses Opus 5.5 and Codex uses GPT-6 Sol unless you pick something else. That's the case on Draftbit credits, with your Claude or ChatGPT plan, and with your own Anthropic, OpenAI, or OpenRouter API key. Threads you've already started keep the model they began with.

Reasoning effort

Opus 5.5, GPT-6 Sol, and GPT-6 Luna all start at medium effort unless you choose otherwise. That's one step below where Opus 5 started. Medium handles most everyday changes well and keeps credit use down. For a stubborn bug or a big refactor, open the Reasoning effort menu next to the model and pick High or above.

Browser controls in the preview

Previews on the canvas now work more like a browser. You get Home, Back and Forward, Reload, and Reset buttons, an address bar you can type a route into, and a searchable route picker in both the single-route and all-routes views. You can also pop a preview out into its own window. Reset takes you back to the starting route and keeps you signed in.

Importing improvements

We launched project import last week, and it's already better. Backend apps can be imported now, and projects with a nonstandard Vite setup import more reliably, since a setup step that fails no longer stops the sandbox from starting. We've also made the migration step required, because skipping it left imported apps in a state Draftbit couldn't work with.

Other Claude Code improvements

We've also shipped a batch of Claude Code fixes over the past few days to make it safer, more reliable, and easier to use. Connecting your Claude plan is easier: you now get 15 minutes to paste the sign-in code, and if it expires you can start a fresh sign-in right where you are. Resumed threads keep following your current instructions, and runs now finish only once all of the agent's work is in, so a failure at the very end doesn't get hidden. We also track context and usage more accurately along the way, so the Context Window Usage indicator stays steady when the agent uses helper subagents and shows the right size after a conversation is compacted.

Custom domains

Setting up a custom domain for your web app is clearer now. The setup dialog used to show about one DNS record at a time, so it was easy to add the first one, click Verify, and hit a failure. Now every record you need is listed up front, grouped by domain and marked required or optional (the optional ones set up the www redirect), with two numbered steps: add the records, then verify. Removing a domain also asks you to confirm first.

Supabase publishing fixes

We've fixed a batch of issues with deploying a Supabase backend to production. Updated edge functions now always make it to production (before, a function that already existed there could be skipped), your auth settings copy over reliably, and if that step fails the whole deploy stops instead of quietly dropping changes.

Your development branch is also reused between deploys now, so you keep your test data, and it's refreshed from production after each deploy. When you want a clean slate, use Fresh dev branch. And if a security check flags something you've set up on purpose, you can review the warning and choose Deploy anyway.

Opus 4.6 retired

We've retired Claude Opus 4.6, so it no longer shows up in the model picker. Opus 4.5 was retired earlier.

Give Opus 5.5 and GPT-6 Sol a try on your next thread, and if anything feels off, reply here or reach out and we'll take a look. Happy building!