[FEATURE] Add support for 1M Context on Claude Sonnet 4.5
Resolved 💬 54 comments Opened Sep 29, 2025 by AntonioIbarraOrtiz Closed Jan 5, 2026
💡 Likely answer: A maintainer (dicksontsai, collaborator)
responded on this thread — see the highlighted reply below.
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Add support for 1M Context on Claude Sonnet 4.5
Proposed Solution
To use the 1M token context window, include the context-1m-2025-08-07 beta header in your API requests
Alternative Solutions
_No response_
Priority
Medium - Would be very helpful
Feature Category
CLI commands and flags
Use Case Example
_No response_
Additional Context
_No response_
54 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
This beta header flag (context-1m-2025-08-07) need to be set on Claude Sonnet 4.5 model.
https://www.anthropic.com/news/1m-context
We are currently limiting access "in public beta on the Claude Developer Platform for customers with Tier 4 and custom rate limits" as well as some Claude Max 20x users.
I have Tier 4
Can you screenshot what you see with
/model? You should see aSonnet (1M Context)option if your organization is Tier 4. Otherwise, you can always configure your model as/model sonnet[1m]. https://docs.claude.com/en/docs/claude-code/model-config#extended-context-with-%5B1m%5DLook like it still not work for Max 20x
⎿ Set model to sonnet[1m] (claude-sonnet-4-20250514[1m])
⎿ API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"The long context
beta is not yet available for this subscription."},"request_id":"req_011CTdp6B5hciTqYYxehWAh6"}
⎿ Set model to Default (Opus 4.1 for up to 50% of usage limits, then use Sonnet 4)
Wait, if it now use Sonnet 4.5 then it have 1M context?
Hmm, look like no, claude code still use claude-sonnet-4-20250514
⎿ Set model to sonnet (claude-sonnet-4-20250514)
Okay, I can specific model, and 4.5 still 200k token
⎿ Set model to claude-sonnet-4-5
⏺ Hello! How can I help you today?
⏺ I'm powered by Claude Sonnet 4.5 (model ID: claude-sonnet-4-5).
⏺ Claude Sonnet 4.5 has a 200,000 token context window.
<img width="1093" height="764" alt="Image" src="https://github.com/user-attachments/assets/6f4ea106-bb28-4497-8643-8f829a514815" />
Is working, I set the model with "/model sonnet[1m]" and now I have 1M context.
Thanks
Not work for 20x subscription
⎿ Set model to sonnet[1m] (claude-sonnet-4-5-20250929[1m])
⎿ API Error: 400
{"type":"error","error":{"type":"invalid_request_error","message":"The long context
beta is not yet available for this
subscription."},"request_id":"req_011CTfbv3FkCcUmY5HZzPnnm"}
⎿ Set model to opus (claude-opus-4-1-20250805)
⏺ Hello! I'm ready to help you with the project. What
would you like to work on today?
We have not rolled out 1M to all Max 20x users yet.
I see, do we have any ETA?
Will I get notification if I am on the list?
Just putting it out there, im a 20x user, willing to test as well :)
Any ETA on when it will be available for Max 20x users?
⎿ API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"The long context
beta is not yet available for this subscription."},"request_id":"req_011CTq5dDpowSKDbT9UGboMG"}
Logout -> login -> still not work
Any ETA on when it could be out for all Max 20x?
What about Max 5x plan user?
Need this ASAP on 20X. Not even sure what there is to gate here, as all this does is use up my usage faster if I need it.
Still not available for 20x users. Even with /model sonnet[1m]
Yep, still not available for 20x without any ETA...
Our team already hit the $5,000 cap for this month on our Tier 4 Team subscription. We were making great use of 4.5 w/ 1M.
In lieu of that, I signed up on my personal account for the Max 20x. I did not realize I would still be limited to a lower context window.
We have tried for months to get a response from Sales on upgrading to an Enterprise account to avoid this. While this may not be the appropriate place to gripe about a lack of response from the Sales teams, I definitely want to toss our team's name in the hat for some kind of beta access to Max 20x w/ 1M.
Same here, had it as 20x but somehow lost access and no longer part of subscription now
I’m a 5x user, and I’m genuinely disappointed. I don’t understand why the older Sonnet model with the 1M-token context window was removed without offering a suitable replacement. If the Claude team doesn’t want to grant access to the newer version, the least they could do is restore access to the older 1M-token model.
max 20x subscribers are supposed to have this access. what the point of have Claude Max 20x then, instead, just more limit. A little bit of incentive for Claude Max 20X would be nice!
please add the 1m context
Extremely frustrating that our Nov 1 date allowed us to resume using the API, and I find that I've been opted out of the 1M context window "test" or whatever.
When you place the 1M claude vs 200k claude next to one another, it's night and day a better product.
I got 2 max x20 plan account, old one cant use the 1m model, new one can use the 1m model...
îts so stupid...
Would love to have 1M context on Claude Web too for projects!
Why? What is the main reasons? I believe the number of Max x20 is not that big given its price.
Why don't you think this is kind of like an insensitive to the max x20 users as beta tester. I'm not sure what is your marketing strategy right now?
I am also a Max 20x user and am dissatisfied with only 200k tokens. I urgently need to use 1Mi and the CLI does not accept it!
<img width="1260" height="332" alt="Image" src="https://github.com/user-attachments/assets/a9e3a86b-6b15-4287-a9da-c71b884ea6b3" />
In my team, we are also lacking a 1M Sonnet. Please add it to the team subscription.
+1 from us as well. All developers have pro max but no 1m window means loading (even token minmaxed) corporate coding standards (reduces a lot of errors) takes up 60%+ of the 200K context window on start. Its really bad. Reduce the CAG pre-load with claude code and get large-repeat generation mistakes and hallucinations. Increase it and you only have 10-20% oc context left before context gets compressed (and the CAG preload is then lost). Its not great, not great at all.
That said, perhaps it is already there and updated and we need to do a thank you instead (at least for CC)?
(also see AntonioIbarraOrtiz's post)
BUT this works only if there is nothing else in your
.claude/settings.local.json
"env": {
"ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-5-20250929[1m]"
},
this
<img width="306" height="348" alt="Image" src="https://github.com/user-attachments/assets/62a41fed-d2a8-4fdb-9b47-88ef3eb4d0fd" />
The only reliable approach we have found with CC 2.0 is a .bashrc with
export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-4-5-20250929[1m]"
SO it might be wider CC 2.0 settings.local.json ENV setting bug.
Can someone else please confirm the same results?
NOTE settings.local.json PUTTING THIS in settings.json does not function (by design)
Just gonna drop by here and leave some positive feedback:
Sonnet 4.5 with the 1 million token context window may be wildly expensive, but _OMG_ it's worth it.
There are _some_ of the much maligned and prognosticated long context performance issues, but they are absolutely made up for by the avoidance of information lost by avoiding compaction. Claude code feels about 1 standard deviation more capable when it doesn't have to re-learn pieces of information 5x throughout the course of a complex debugging or implementation session, and I say this after pushing subagent orchestration and delegation with the 200k context window about as far as I can go.
This is all even though the majority of my usage of
sonnet[1m]requires less than 500k tokens, as avoiding 2-3 compaction also avoids the need to re-generate like 60-80k of context to reason and plan properly.Who do i have to bribe to get access to this? i have been a claude user for a long time now and i am on the 20x plan.
@BainViljoen
You should not need to, setting this in .bashrc and Claude Max 20x works fine for me.
You can't rely on ANY LLM advice (claude, GPT, etc) on this topic, ironically, it always gets it wrong. I had to nut this one out myself, including the settings.local.json ENV not working when combined. But if this does not work for you still then yep, its probably region, account type, something Anthropic account side.
echo 'export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-4-5-20250929[1m]"' >> ~/.bashrc@hsderek , you almost had me exited, but unfortuantly i am blocked on an API level:
API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"The long context beta is not yet available for this subscription."},"request_id":"req_011CVJKTtwTL3V9ZLKLhbKJq"}
And i so despretly need the large context as my codebase spans across mutiple platfroms that all communicate together.
The 200k context just keeps compacting to soon...
@dicksontsai Who do i have to bribe to get this?
@BainViljoen - if if makes you feel better then I should have said nothing. Now I don't have claude-sonnet-4-5-20250929[1m] access either. Seems they 'fixed' my approach or access in CC instead of fixing adding 1m as an option. :-(
So +1 Please on this one from me too!. CC's 200K window is just too painful. TBH ~500K would do the trick for most of the use cases we see.
Actually, the 1M is not stable as conversation around on the internet. That is why it still in beta now, it is not because they do not make it available for us but to be safe to use.
@silver-epsilo
We've had 1m sonnet 4.5 stable fine for weeks now (until it just got turned off) with claude code and some seriously demanding coding tasks. Plus repeat tests as we have a meta layer over CC we use as well. Usually with restart to a new session before 700-800K tokens.
Where did you find these conversations so I can contribute to them?
The move back to 200K has seriously degraded the effectiveness of Claude Code for us as the compress context at nearing 200K re-introduces constant repeated mistakes that then waste significant human fix up time. And adding the instructions for these back into CLAUDE.md then just consumes more context making the time before a compress increasingly smaller creating a death spiral. Also the compress does not re-read the CLAUDE.md so all the tasks it then immediately commences with post-compress make all the same bad mistakes the last session was prompted not to.
The move from 1m -> 200K has defined tasks that we used to use claude code for that we simply cannot now and have to go back to our own agentic coders which is not as efficient nor as easy to work with or integrated.
Its just plain bad.
A move back to Cursor (+ Gemini 3.0 Pro 1m ) might have to happen, which would be sad as we have a lot of CC integrations built.
Any updates? Would really love the 1M model
Update time>
With the Opus 4.5 update as default the problem is still there and very painful. The number of bad mistakes post compact with Claude Code are huge. There isn't even a 'read this after compact' option. On average we waste around 5 mins reverting mistakes that are made after every compact when it loses context and makes repeat mistakes, usually for the very same things that have been corrected 5x or more previously. This did not occur with ~500K tokens even with sonnet, and definitely not with a 1m limit. We usually manually compact to head off context rot around 600K. I don't buy the official Anthropic "it's for stability" explanation looking at the CC npm and activity, at least not as the primary reason. The primary reason will be cost.
The only reliable way to get 1M right now (I pay for the 20x Max plan as well and don't have access through my subscription) is to pay out of pocket for its use on AWS Bedrock (there you don't need to be "Tier 4" to get access and it's priced the same anyways). It's expensive though. You will easily run up 1k+ bills if you use it as your daily driver.
I am a new CC subscriber (for a few days now). I just launched claude and saw this notice:
So maybe it is back?
I'm a long standing 20x Max ($200/month) subscriber and I still don't have access to the 1M Context model in Claude Code's
/model. If I manually set the model using/model sonnet[1m]it says:API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"The long context beta is not yet available for this subscription."},"request_id":"req_011CWptvePXLjZBoWiVJCE9F"}So new subscribers have it. Longer standing subscribers don't :/. Seems very random.
1m is still arbitrary and opaque for users, it can change on no notice I had 1m before as you can see. We have to assume its not there in operations and if it is then great, make hay while the sun shines. We've moved the agentic coding that really struggles at 200K limit to Gemini for a while though the two stacks are a pain. We will just have to see what Anthropic's plans are. Considering how cutthroat this particular AI arms race is I don't expect the customer will know until it happens. It would be good to know if its even on the cards and if they come back with 'context window is less important' (a Microsoft style reality disconnect response) they will likely get a big backlash.
Wait why was this closed? Is this completed? Everyone has 1M access now?
I don't have
I agree @alisoltani - why is this closed? It'd not fixed. Was this an agentic driven issue review?
So, I have to say this, as unbelievable as it is.
I have forever been an advocate of claude opus but recently my projects grew to large and the limited context window became a problem an unfortuantly it wasnt because the tasks were to big. The solution consists of 4 diffirent systems that all work together so the context window just cant handle a single command some times.
This made me give codex gpt 5.2 xhigh a shot, boy o boy....
It is finding issues that opus never found, and the issues are valid, i have told opus to fix them only for it to make a buggy patch.
GPT on the other hand... its finding all the root causes without me even pointitng to them.
The 400k (272k input and 128k output) goes a long way as well with my multi level project.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.