Claude 3.7 Sonnet [Public Preview]

[matthewisabel]

Claude 3.7 Sonnet is now available in Copilot

See the changelog for Claude 3.7 Sonnet

The new Claude 3.7 Sonnet is now available in Pro, Business, and Enterprise plans across Copilot Chat across Visual Studio Code, Copilot Chat on GitHub.com, and Visual Studio. The model will also be available soon in Jetbrains IDEs.

We'd love to hear your suggestions for improvements and any issues we might see as you start using the new model!

_ 🟢 2025-02-25, 20:45:00 UTC After intermittent instability, we have now rolled out access to all paid Copilot plans._

[hmorneau]

It's out already: https://www.anthropic.com/news/claude-3-7-sonnet

I just tested on Cursor, it's much better.

[matthewisabel]

Cool, hope you'll give it a try on on Copilot as well :)

[hmorneau]

@matthewisabel I just did, and it outperformed Cursor in the first try. Seems like you deployed a good version of it. Thanks!
How can we get rid of the old model like 4o, o1, o3-mini from the list? ;)

[RebaiFedi]

When will it be available on GitHub Copilot

[ericktreichel]

come on guys.. just read two or three comments below... there are 10 comments about they removing 3.7 due to issues. It will be back soon

[spacesnotabs]

Yes, but according to the GitHub Status page, the issue has been resolved, which is why people are still bringing it up.

[ericktreichel]

Yes, but according to the GitHub Status page, the issue has been resolved, which is why people are still bringing it up.

yeah, that "Resolved" status in the status page is very confusing, but people already talked about the meaning of that here 10x as well

[matthewisabel]

Sorry for the confusion today folks, now that we're fully back up with the new model please let us know if you hit any issues.

[sp-aaflalo]

@matthewisabel Are you sure it's fully back up ?
I still don't see it, even when enabled on our organization Copilot Models and Policies. I only see the Claude 3.5 and Gemini 2.0 Flash in any UI, on github or vs code. I only see it in Jetbrains IDE oO

[tzutoo]

Cursor, Cody, Roo code have announced support for Sonnet 3.7. Windsurf is expected to support it soon. Hope GitHub copilot supports it as well.

[RebaiFedi]

But GitHub Copilot is slow in taking updates..

[hmorneau]

Not today, it's available now! Copilot is really rocking these day. I don't know what happened but they are getting good now.

[lxyesa]

I can now see that the model is available for selection in VSCode, but it is currently not usable and has returned an error: Request Failed: 400 {"error":{"message":"The requested model is not supported.","param":"model","code":"model_not_supported","type":"invalid_request_error"}}

[CRooi]

same here

[kkjdaniel]

Yup model is completely gone for me too.

[razvanphp]

This incident has been resolved.
Posted 2 hours ago. Feb 25, 2025 - 15:45 UTC

Still not shown in my VSCode 🙁

[matthewisabel]

Sorry for the confusion today folks, now that we're fully back up with the new model please let us know if you hit any issues and what you think.

[poetic-design-net]

do you will open it in also for VS Code LM API? @matthewisabel

[muzzlol]

i get the following prompt repeated back to me.

forgive me if im being ignorant, but shouldnt updating the model essentialy be a one liner change? just trying to understand how things work i appreacite everyones effort

[matthewisabel]

Hey are you still seeing this? It is possible this happened during rollout based on timing.

[muzzlol]

Hey are you still seeing this? It is possible this happened during rollout based on timing.

no, everything working perfectly good sir
👍

[hmorneau]

@matthewisabel I was able to output 6000 tokens and it wasn't even the limit. 🥇

Edit: well, that didn't last long, see below.

[hmorneau]

@matthewisabel They nerfed Claude 3.7 thinking already, that was fast. The exact same prompt last night as reported above, I was able to convert that document and get an answer of 6200 tokens.

Now today on a new conversation, same prompt, the same file show that's it's over the context limit (it's in orange).

The answer yesterday was 6000 tokens. The answer today 3500 tokens and it's obviously incomplete.

It didn't last long...

[wuaimtuangnueia]

Does sonnet 3.7 (and think-mode) have the quota of using? Is it the same unlimited as 3.5?

[muzzlol]

idk about thhinking, but 3.6 sonnet and 3.7 have same pricing so they probably would'nt rate limit this if it wasn't already

[wuaimtuangnueia]

idk about thhinking, but 3.6 sonnet and 3.7 have same pricing so they probably would'nt rate limit this if it wasn't already

Usage limits are not necessarily set based on price. For example, the o3mini should actually be cheaper than the 4o, but the o3mini is limited to 50 times per 12 hours in copilot.

[matthewisabel]

We have a rate limit for 3.7 that's a bit lower than 3.5 presently as we ensure rollout reaches all users with plans to raise it over the week to 3.5 levels.

[wuaimtuangnueia]

We have a rate limit for 3.7 that's a bit lower than 3.5 presently as we ensure rollout reaches all users with plans to raise it over the week to 3.5 levels.

Will o3mini remove the 50-per-12-hour limit? As mentioned above, o3mini actually costs less than 4o

[pidgeon777]

We have a rate limit for 3.7 that's a bit lower than 3.5 presently as we ensure rollout reaches all users with plans to raise it over the week to 3.5 levels.

Hi @matthewisabel, did you mean over the week(s)? Or the 's' is intentionally missing (I hope so 🙂).

[ericc-ch]

Are the thinking/reasoning tokens hidden for the thinking model? I can't seem to find it

[matthewisabel]

We don't show them at the moment but will soon

[ericc-ch]

Thank you for the answer! Nice to hear that, hope it will be shown soon

[devslimbr]

Using Claude Sonnet 3.7, I can barely use it properly. I keep getting rate limited, and now I’m blocked for 2 hours, even on the paid version

[matthewisabel]

We're seeing some error rate elevation with AWS Bedrock and capacity constraints at the moment, we're working actively to resolve

[linacloud-fairy]

When will vision be incorporated into the model?

[matthewisabel]

In the near future we'll add this model into our vision preview

[giovannicocco]

Hi! I am getting 500 Internal Server Error and sometimes code 400. I got this issue using Copilot Edits, Cline and Roo Code from VS Code and VS Code Insiders

[LBV2012-26]

Yes, very unstable.

[mmw1984]

[hmorneau]

@mmw1984 and all that yellow on your css and js file mean you are over the context limit.

[mmw1984]

@mmw1984 and all that yellow on your css and js file mean you are over the context limit.

known

[indcoder]

Getting the 500 errors on more than one occassion

[matthewisabel]

This should now be resolved but feel free to reach out if you continue to see issues

[indcoder]

Yes, thank you @matthewisabel for letting me know. For all the 4 prompts I had, I did not encounter any issues.

[ricsi1012]

Is Claude 3.7 Sonnet available in agent mode? I can only see it in 'edit' mode.

[giovannicocco]

I think this morning 4-5am (CT)

[mmw1984]

I think this morning 4-5am (CT)

woww...... thats why I can't find now, No reason why they remove them.

[hmorneau]

I think this morning 4-5am (CT)

woww...... thats why I can't find now, No reason why they remove them.

Still there for me, I guess as long as I don't switch away from it. Just tested and it's working, so much be a UI thing only.

[mmw1984]

working, so much be a UI thing only.

wish its a bug

[matthewisabel]

Sorry for the frustration here with some early instability. You should see 3.7 in Agents/Edits and we're working on enabling the thinking model here as well.

[mmw1984]

Please add Claude 3.7 Sonnet back to Agent mode in VS Code

[owen800q]

Please add Claude 3.7 Sonnet back to Agent mode in VS Code

佢加左，你update下vs code

[mmw1984]

真??! Insider... 再試下update..

Eng Trans: Really??? I am in Insider version, let me try to update again

[mmw1984]

[matthewisabel]

Sorry for the instability today folks, now that we're fully back up with the new model please let us know if you hit any issues and what you think.

[comontes]

🥹 loving the new model

[tleo19]

getting a lot of: Sorry, no response was returned

[waleed1892]

Getting rate limit even on first message. Currently sonnet 3.7 is unusable with github copilot.

[waleed1892]

for me it's continuously giving rate limit error.

[matthewisabel]

We had lower rate limits this morning but have since increased. They're still hittable with a lot of agent mode usage but a good bit above what they were.

[matthewisabel]

If you're hitting continuously let us know.

[waleed1892]

limits seems to be generous now but getting this error in every request and even with multiple retries this popup appears. After this error copilot mess up previously written code.

[Wituz]

I also got rate limited with Sonnet 3.7, and should wait for 87 minutes

[cmp-nct]

"Sorry, the response hit the length limit. Please rephrase your prompt."

It's useless

[WyvernWin]

It's annoyingly useful which makes it utterly useless right now haha - at times when I guess lower network traffic was happening it was working great in agent mode, whirling away fixing stuff - enhancing stuff .... Then it's not available and suddenly I see it's tried to agentically build visuals into realtime batch logging and epoch analysis through a dash web server - all locally, let's just say machine learning and the memory leak into cpu and allsorts of lovely stuff I have to rollback or untangle - rollback lol

IF it was available in full context I do feel Claude was gonna make it work even when confronted with using tensor board it said, nah watch me..."connection lost"

[RomDlg]

Claude 3.7 sonent is actually unusable for me, im getting rate limited even on first message

[Liad0205]

The new model it awesome! Is there a chance Grok3 thinking or o3high will make it too?

[matthewisabel]

Hopefully, we'll see

[resslr]

Anytime I try to use the model with the #codebase flag I get this error Unknown tool: "semantic_search"

[mary-ext]

Now that we have several reasoning models available like o1, o3-mini, and Claude 3.7 Sonnet Thinking, is there any plan for showing the thinking tokens in chat?

The thinking tokens are very useful as a debug mechanism for prompts.

[matthewisabel]

We plan to in the near future!

[BoldPlut0]

I'm using claude code with aws bedrock. Is anyone getting ratelimiting errors just after the 1st message ?

summarize this project for me 
  ⎿  API Error (429 Too many tokens, please wait before trying again.) · Retrying in 16 seconds… (attempt 6/10)
  ⎿  API Error (429 Too many tokens, please wait before trying again.) · Retrying in 1 seconds… (attempt 1/10)
  ⎿  API Error (429 Too many tokens, please wait before trying again.) · Retrying in 1 seconds… (attempt 2/10)
  

[matthewisabel]

If you're not using GitHub Copilot you may want to go to Bedrock forums for support or Anthropic for Claude Code

[kiwina]

Hard pass for now 5min usage and locked me out for 3+ hours never seen that long wait time. it seems better but i don't want to read a book when i refactor code, basically every iteration in agents it spits out more useless repetitive comments than actual code.

[matthewisabel]

That's surprising you could hit a rate limit in 5 minute but agent mode can use a lot of tokens. We've raised the rate limits since rollout but a lot of sustained usage can still hit limits.

[denysgaievskyi]

My company is using copilot business plan, 3.7 is enabled by admin, but nobody from the team can see the new model in the dropdown menu(JetBrains plugin). what can be the issue?

[lurenjia534]

Currently JetBrains plugin does not support Claude 3.7 in the model selector. They are always slow to support JetBrains plugin (compared to vscode)

[denysgaievskyi]

thanks!

[lurenjia534]

Hi I just noticed that JetBrains plugin now supports Claude 3.7. I have seen it. Update your copilot plugin in idea

[Josexv1]

Works good but dang you hit the limit really fast, when i say fast i used the think mode to refactor a single file around 700 lines, asked to add another thing, then i asked to check a folder to see why an endpoint was not protected and 3 hours of rate-limit, not even o3 have these hard hitting limits. Back to 3.5 i guess...

[wysiecki]

i hit the limit on every request now

[matthewisabel]

Hm, we'll look into it. Do you mind sharing the exact message you see? Is it after sustained usage?

[wysiecki]

Just started after break, first task, it made some requests, and then:

Sorry, your request was rate-limited. Please wait 34 seconds before trying again.
on "continue"
Sorry, your request was rate-limited. Please wait 3 minutes before trying again.

So the time does mean not much, i guess.

Thank you, sir

[demiro]

I don't know about you guys... but the same thing happened to me with sonnet 3.7 than it did with o3mini... First day was working marvels.. everything on the first try ... since then, it degraded massively

also to me, it happens many times, that the model just deletes a huge chunk of my code - just deletes it ...
whilest claude is polite about it when confronted... o3 just keeps denying it...

not sure if this happens after it loses context or on the first try (cannot seem to find the way in vscode to find the history of the interactions in edits tab) - or it loses context after going back and forth.... usually my flow is I give it prompt and feed the whole codebase for context [which is a small nuxt project - only shadcn components are many] - and maybe going back and forth ... only after 2-3 sub-prompts... I need to feed the codebase again, because it loses it quick....

so my thoughts - THIS SHOULD NEVER HAPPEN - that the model suggest deletes of the code that are very important (even if not) and is unwarranted .. given other task...

second... I get the models have limited context window.... but we need to find a better solution to keep whole codebase in context - at least a way to automatically identify the most important/pertinent part of it...

[a-bran]

Yeah o3-mini behaves very badly at tool-calling and claude can quickly become overwhelmed without proper guideance.
For context-window an outline-generator like aider uses might be a good solution.

Copilot is nice, but Cline/Roo Code is still better and more autonomous, but is also still missing some features like reading/editing multiple files at once, a code outline to save context, and it has to scale down request- and token-usage massively

[pidgeon777]

I've a question about Claude Sonnet 3.7 input tokens.

If I recall correctly, with 3.5 we have 200k tokens per input (context window). What about 3.7?