Anthropic shipped Claude Sonnet 5 on June 30, 2026. On paper, it is another mid tier model release. In practice, it is the first Sonnet class model that behaves like a genuinely agentic coding partner, and it is already the default in Claude Code, in most Lovable projects, and inside a growing list of IDEs. If you build software by describing it rather than typing it, this release matters more than the version bump suggests.
Here is the honest, grounded read.
The headline numbers
Sonnet 5 posts 63.2% on SWE-bench Pro, the agentic coding benchmark that mirrors real developer tickets more closely than the older SWE-bench Verified set. That is a meaningful jump from Sonnet 4.6 and it lands at the same $3 in, $15 out per million token pricing. Anthropic also reports big drops in step count on long tasks, which is the number that actually shows up on your credit bill in a tool like Lovable.
Translation. Same price, fewer round trips, better outputs on the tasks that look like real work. Not a lab benchmark win, a workflow win.
What actually feels different
I have been running Sonnet 5 through the same test bed I use for every new model: a landing page, a small SaaS dashboard with auth, a checkout flow, and a bug fix on an existing codebase. Three things stand out.
First, it plans before it types. On a request like "add Stripe checkout to this project," older models would jump straight to writing a component. Sonnet 5 usually stops, reads the existing routes, and proposes an approach before touching a file. That maps directly to how Plan mode works inside Lovable, and it means fewer wasted Build cycles.
Second, it stops looping. The failure mode of previous Sonnet models on hard bugs was to try the same fix four times in a row. Sonnet 5 is meaningfully better at recognizing that it is stuck and asking a clarifying question or trying a different angle. That alone will save credits for anyone who has watched an agent burn through ten prompts on the same TypeScript error.
Third, it reads the whole file. Long context handling is noticeably better on projects with fifty or a hundred files. Older Sonnet models would sometimes edit the wrong function because they missed a similar name elsewhere in the codebase. Sonnet 5 does that less. Not never, but less.
Where you will feel it first
- Lovable projects. Sonnet 5 rolled out as the default coding model for most Lovable accounts within about a week of launch. If your projects feel snappier and your credit usage per prompt looks lower, this is why. The pricing breakdown covers how credits are counted.
- Claude Code. Anthropic's terminal based coding agent got the biggest lift. Long tasks like migrating a codebase or refactoring a service now complete in far fewer steps. See the full Claude Code vs Cursor vs Copilot comparison for context.
- Cursor and other IDEs. If you use Cursor with a Claude key, switch the coding model to Sonnet 5 and give it a week before deciding. It benefits most on multi file changes.
- Agents you built yourself. If you wired the Anthropic API into a custom agent, the drop in step count is the number to watch. Anthropic reports 25% fewer steps on typical agentic loops, sometimes more.
Prompting habits that get more out of it
- Give it the constraint, not the solution. Older models needed hand holding. Sonnet 5 does better when you describe the problem and let it choose the shape. "Add a booking flow that saves to Supabase, sends a confirmation email, and shows a success screen" is a better prompt than a 300 word spec of files to touch.
- Ask for a plan first on anything non trivial. Even outside Lovable, "before you write code, walk me through your plan in five bullets" is a prompt worth pasting once per session. Sonnet 5 will actually stick to the plan you approve.
- Stop rewarding loops. If the agent tries the same thing twice, cancel and give it new information. It is smart enough now that repeating yourself is a signal to change tactics, not to try harder.
- Use the Knowledge Base. In Lovable, a half page Knowledge Base plus Sonnet 5 is the combo that shortens the most prompts. It stops repeating itself, and it applies your tone and constraints automatically.
The honest limits
Sonnet 5 is not Opus. On very hard reasoning tasks, on very long custom protocols, and on some math heavy code, you will still want the top tier model. Anthropic positioned Sonnet 5 as the cheap way to run agents at scale, and that is exactly what it is. For a solo builder shipping on Lovable, that is the sweet spot most of the time.
It is also not a magic security review. It will happily wire an insecure auth flow if you ask for one. The beginner mistakes guide still applies in full. Better models raise the ceiling of what one prompt can do, they do not remove the responsibility of reviewing the code that ships.
Should you switch anything?
If you are on Lovable Pro, nothing to do. The default already routes to Sonnet 5 for most requests. Reap the credit savings, ship more.
If you use Cursor or Claude Code, update your model setting and pay attention over the next week. If you use another Claude powered tool, check what version they are on. If they are still on 4.6, ping them.
If you have never tried Claude directly and want to see the model behavior outside a builder, the free week of Claude is the easiest way to feel the difference on your own workflow. It also helps me out with usage credits when you subscribe, so honest disclosure.
The bigger picture
Every twelve months, the ceiling of what one prompt can do goes up. Sonnet 5 is the clearest step change of the year on the coding side. Combined with Lovable's agent integrations and the platform's fresh round of funding, the practical distance between an idea and a running app keeps shrinking.
The people who benefit most are the ones actively shipping when the model lands. If you have been meaning to start, the model is not the bottleneck anymore. Open Lovable, describe something small, and see what one Sonnet 5 backed prompt can do this week.
