Context Window

[ryukenshin546-a11y]

Select Topic Area

Question

Copilot Feature Area

VS Code

Body

Why does the Claude Opus4.6 token context window only have 128K input and 64K output, when the model can handle up to 1M?

[ryukenshin546-a11y]

[aryankumar06]

Claude Opus 4.6 “supporting 1M context” is basically the model’s maximum architectural capability, but in the actual product Anthropic sets practical limits for speed, cost, and reliability. At 1M tokens the compute and latency blow up massively, and even if the model can technically read that much, its effective recall and consistency can become less stable. That’s why they cap it at 128K input and 64K output to keep performance predictable, inference fast enough, and pricing manageable.

[anuragchauhan06]

In cursor you have 200k token context window and you're absolutely right.

[SeaDude]

I believe the 128k limit is imposed by GitHub Copilot's agent harness. Just like Claude Code, Cline, Cursor, etc., under the hood, GitHub Copilot is a "coding assistant agent" with a "harness"; a combination of programmatic and non-deterministic loops that handle input, context, and output.

If you want to work with 4.6's 1M context window right now, you'll have to select an agentic tool whose harness can handle 1M OR create your own custom harness.

[MuhammedSinanHQ]

Claude Opus 4.6 can support up to ~1M tokens at the model level.

Copilot Chat does not expose the model’s maximum context.

Copilot applies its own caps for:

• Cost control
• Latency
• Reliability
• Tool orchestration
• Multi-tenant fairness

So you see:

128K input / 64K output
even though the underlying model supports more.

---

Important distinction

Model capability ≠ Product limit.

Vendors frequently gate large contexts behind:

• Enterprise tiers
• Private preview
• Custom contracts

Copilot currently uses a constrained configuration of Claude models.

---

Why Copilot does this

• Long-context inference is expensive
• Tool calls + embeddings scale with context
• UI responsiveness degrades past ~100K tokens

GitHub optimizes for interactive developer workflows, not massive document ingestion.

---

If you truly need >128K

Use Claude directly through Anthropic’s API or a provider that exposes 1M context.

Copilot is not the right surface today.

[nickchomey]

Claude Opus and Sonnet 4.6 both just changed 1M context window from beta to GA. Copilot should make use of it as well

https://claude.com/blog/1m-context-ga

[Hronom]

When this will be available in GitHub Copilot? Or we need to migrate to Claude Code?

[raaf003]

pretty disappointed with context window for Claude opus 4.6 in copilot.... its not even usable as it easily reaches limit .... compared to open and gemini modals have huge context windows

[wiltaylor-daf]

Over the weekend got to play with Claude Code and the 1M token window. The difference is night and day. Lets you do some really complex work without having to worry about the main agent getting anywhere near the context window. Keeps long running agentic work on track without the risk of compaction causing it to go off track. We really need this in copilot asap.

[ryanjordan11]

Yeah more memory, more bullshit.
They hype the million-token flex like it’s some god-tier upgrade, but then they choke it down to 128K because.. what? Latency spikes? Costs explode? Or just honestlyn. because nobody wants to wait ten seconds for a reply while the model chews through your whole codebase.
It’s all marketing. "Look! We can do 1M!" Cool. Now do it without making me pay extra or sit there like a fool.Truth? The real limit isn’t the tokens it’s the humans running the show. They’d rather you stay small, stay fast, stay hooked. More memory just means more ways to screw you over.

[mGaosi]

How about upgrading to 400k？I tested GPT5.4-400k and found that context compression reduces some of it in my workspace. Of course, it may not be possible to make such a simple comparison.

[niquedegraaff]

Maybe 1M is too big, but 200k is too small. Middle ground would be 400k for me. Letting it work on a task without "compacting conversation and ai model getting confused right after and messing the whole progress up..".. well that would be great.

[EdLovecraft]

With such a small context window, Opus looks like a complete idiot when handling complex projects.

[Gecko51]

GitHub Copilot sets its own context limits independent of what the underlying model supports. The 128K input / 64K output cap on Opus 4.6 is a GitHub product decision, not something Anthropic controls.

Why GitHub does this: running 1M token inference at scale is expensive and significantly slower. Interactive coding workflows are latency-sensitive, and serving 1M contexts across a large user base with sub-second response expectations doesn't work economically right now. So GitHub capped it at something usable for the product.

The 1M context window did go GA on the Anthropic side in early 2026, but GitHub hasn't updated Copilot's limits to reflect that yet. There's no public timeline for when (or if) they'll increase it.

If you need the full 1M window, your options are Claude Code or the Anthropic API directly. Copilot is not the right surface for that use case at the moment.