github-copilot: add 1M-context (`-1m`) Claude model variants

[xaviergmail]

Summary

On 2026-06-04 GitHub announced that Copilot now supports 1M-token context windows and configurable reasoning levels across VS Code, Copilot CLI, and the GitHub Copilot app, "expanding to more surfaces soon":

📣 https://github.blog/changelog/2026-06-04-larger-context-windows-and-configurable-reasoning-levels-for-github-copilot/

Copilot exposes these capabilities via separate model IDs on the /models CAPI endpoint — model IDs with a -1m suffix, distinct entries from the 200k-context base models. The github-copilot provider on models.dev currently lists only the 200k base variants. The 1M variants are missing.

This matters because third-party clients use models.dev as their catalog. For example, opencode's opencode models --refresh literally fetches models.dev/api.json, so users cannot select the 1M variants even though the Copilot API serves them.

Related issues

• GitHub Copilot provider missing Claude Opus 4.6 (1M context) model #1292 (open) — requests adding claude-opus-4.6-1m. Pre-dates yesterday's announcement; the announcement now makes this much more pressing and broader in scope.
• GitHub Copilot models report base-model context (~1M) instead of Copilot's actual CAPI limits #1946 (closed 2026-06-02) — established the canonical pattern for github-copilot: model entries override [limit] with CAPI-actual values instead of inheriting from the upstream base model. This issue extends that pattern by adding the missing 1M variants as separate model files.

Missing model variants

Cross-referenced against CAPI dumps cited in #1292 comments and public reverse-engineering writeups (openclaw/openclaw#72824):

CAPI model ID	Context	Output	Reasoning (effort)	Status
claude-opus-4.6-1m	1,000,000	64,000	low / medium / high	GA — exposed to OAuth tokens
claude-sonnet-4.6-1m	1,000,000	64,000	—	GA
claude-opus-4.7-1m-internal	1,000,000	64,000	low / medium / high / xhigh	The -internal suffix in the CAPI model ID strongly suggests this is still first-party-only (VS Code / Copilot CLI / Copilot app). Recommend holding until GitHub drops the suffix, OR adding with a preview = true / experimental marker.

Are there other token-count variants?

Checked the obvious size/capability suffixes (-128k, -32k, -500k, -long, -extended, -fast, -turbo, -vision) in:

• The official docs page at https://docs.github.com/en/copilot/reference/ai-models/supported-models
• Recent dumps of /models from public repositories
• The microsoft/vscode-copilot-chat model registry

Conclusion: as of 2026-06-05, only -1m size suffix variants exist on Copilot CAPI, and only on the three Claude models above. GPT-5.x and Gemini families currently have a single base entry on Copilot (per #1946 verified limits: gpt-5.5=400k, gemini-3.1-pro-preview=200k). No other context-size variants are exposed. Maintainers should re-verify periodically since the announcement explicitly says "expanding to more surfaces soon".

How to verify with a live CAPI dump

Same one-liner from #1946, filtered for size-suffix model IDs:

curl -s https://api.githubcopilot.com/models \
  -H "Authorization: Bearer <copilot-token>" \
  -H "Editor-Version: vscode/1.0" \
  -H "Copilot-Integration-Id: vscode-chat" \
| jq '.data[] | select(.id|test("-1m|-long|-extended|-fast|-turbo|-vision")) |
      {id, ctx: .capabilities.limits.max_context_window_tokens,
            prompt: .capabilities.limits.max_prompt_tokens,
            out: .capabilities.limits.max_output_tokens,
            effort: .capabilities.supports.reasoning_effort}'

Suggested file additions

Mirror the post-#1946 shape. New files under providers/github-copilot/models/:

claude-opus-4.6-1m.toml

base_model = "anthropic/claude-opus-4-6"

# Pricing TBD — Copilot billing for the 1M variant may follow Anthropic's
# "long context" 2x multiplier above 200k input tokens, or it may follow
# the standard Copilot premium-request multiplier. Please verify against
# https://docs.github.com/en/copilot/managing-copilot/managing-copilot-as-an-individual-subscriber/about-billing-for-github-copilot

[limit]
context = 1_000_000
input   = 800_000   # verify via CAPI max_prompt_tokens
output  = 64_000

claude-sonnet-4.6-1m.toml

base_model = "anthropic/claude-sonnet-4-6"

[limit]
context = 1_000_000
input   = 800_000   # verify via CAPI max_prompt_tokens
output  = 64_000

claude-opus-4.7-1m-internal.toml — hold until GA (CAPI ID still has -internal suffix), OR add with a preview flag.

Also flagged: configurable reasoning levels

The same June 4 announcement covers "configurable reasoning levels". The upstream providers/anthropic/models/claude-opus-4-7.toml already declares:

[[reasoning_options]]
type = "effort"
values = ["low", "medium", "high", "xhigh", "max"]

…but the github-copilot models inherit only [cost] / [limit] overrides and may not propagate reasoning_options correctly. Worth verifying whether downstream clients see the effort tiers when selecting a github-copilot/claude-opus-4.x model, or whether this needs an explicit declaration. Likely a separate issue if it turns out to be broken.

---

Filed by an opencode user who hit the 200k context limit on github-copilot/claude-opus-4.7 immediately after seeing yesterday's announcement. Happy to open a PR with the two GA files if a maintainer confirms the suggested shape (especially the pricing/billing piece).

[xaviergmail]

Correction / expansion — after additional research I have to retract my "only -1m variants exist on certain Claude models" claim. The community plugin opencode-copilot-1m (npm) registers 6 separate -1m variants, including GPT-5 family models I had previously ruled out. Microsoft's first-party VS Code source also distinguishes a non-internal claude-opus-4.7-1m from claude-opus-4.7-1m-internal.

Additional missing variants

CAPI model ID	Context	Input	Output	Source
claude-opus-4.7-1m (non-internal)	1,000,000	900,000	32,000	vscode toolSearchTool.ts lists it as a distinct family from -1m-internal; opencode-copilot-1m/index.ts registers it client-side
claude-opus-4.8-1m	1,000,000	900,000	64,000	opencode-copilot-1m/index.ts
claude-sonnet-4.6-thinking-1m	1,000,000	—	64,000	Combines thinking-mode + 1M; appears in third-party catalogs (Windsurf / OmniRoute)
gpt-5.4-1m	1,050,000	922,000	128,000	opencode-copilot-1m/index.ts
gpt-5.5-1m	1,050,000	922,000	128,000	opencode-copilot-1m/index.ts

Note GPT-5 1M variants use 1,050,000 (not 1,000,000) and a different input cap. They also have tiered pricing that kicks in above 272,000 tokens — the plugin captures this:

// gpt-5.4-1m
cost: {
  input: 2.5, output: 15, cache_read: 0.25,
  tiers: [{ type: "context", size: 272000, input: 5, output: 22.5, cache_read: 0.5 }]
}

// gpt-5.5-1m
cost: {
  input: 5, output: 30, cache_read: 0.5,
  tiers: [{ type: "context", size: 272000, input: 10, output: 45, cache_read: 1 }]
}

Above 272k tokens, input cost 2x and output cost 1.5x. models.dev's TOML schema appears to support [[cost.tiers]] based on existing entries elsewhere — worth confirming.

Updated suggested TOML additions

In addition to the three files in the original issue, add:

claude-opus-4.7-1m.toml (note: the api.id should be claude-opus-4.7-1m, NOT claude-opus-4.7 — they're separate CAPI model IDs):

base_model = "anthropic/claude-opus-4-7"
[limit]
context = 1_000_000
input   = 900_000
output  = 64_000

claude-opus-4.8-1m.toml:

base_model = "anthropic/claude-opus-4-8"
[limit]
context = 1_000_000
input   = 900_000
output  = 64_000

gpt-5.4-1m.toml:

base_model = "openai/gpt-5.4"
[limit]
context = 1_050_000
input   = 922_000
output  = 128_000
[cost]
input = 2.5
output = 15
cache_read = 0.25
[[cost.tiers]]
type = "context"
size = 272_000
input = 5
output = 22.5
cache_read = 0.5

gpt-5.5-1m.toml:

base_model = "openai/gpt-5.5"
[limit]
context = 1_050_000
input   = 922_000
output  = 128_000
[cost]
input = 5
output = 30
cache_read = 0.5
[[cost.tiers]]
type = "context"
size = 272_000
input = 10
output = 45
cache_read = 1

Required HTTP header (for downstream consumers / docs)

The opencode-copilot-1m plugin also pins a header on every request:

Copilot-Integration-Id: copilot-developer-cli

Without this header, the Copilot CAPI may reject or downgrade requests to the -1m variants. This isn't a models.dev concern strictly (it's per-client), but worth noting in the github-copilot provider.toml docs or as a [headers] block if the schema supports per-provider required headers.

What I'm still NOT certain about

• The reasoning-effort-tier variants for opus 4.7 (-high, -xhigh, -max, -review) that appear in third-party catalogs (Windsurf, OmniRoute) — these may be client-side virtual models layered on top of the base CAPI ID by routing tools, NOT actual CAPI model IDs. I would NOT add these without a live /models dump confirming.
• Whether claude-opus-4.6-fast / claude-opus-4.8-fast are CAPI model IDs or feature flags on the existing base models. The existing TOMLs handle "fast mode" via [experimental.modes.fast] blocks with the speed = "fast" body parameter, which suggests it's a parameter not a model ID.
• gemini-3.1-pro-preview-1m — plausible given the 1M trend but no concrete source found.

Sorry for the back-and-forth in one thread — the 1M Copilot ecosystem is moving fast and there's a lot of derivative-catalog noise (Windsurf / OmniRoute / hermes-agent) that mixes CAPI IDs with their own routing-layer abstractions. A live CAPI dump from a maintainer is the only authoritative way to disambiguate.

[yukoba]

As of June 8, 2026, the information in Visual Studio Code shows that the context size is 1M even without appending -1m.

Also, MAI-Code-1-Flash (mai-code-1-flash-picker) has been added.

In #1934, I provided the ID for Raptor mini as raptor-mini, but it seems the correct ID is actually oswe-vscode-prime@raptor-mini.

[xaviergmail]

Shows as 1M in opencode today as well, with no change made on my part, I think it's safe to close this.

[MojangPlsFix]

Seems like the Opus Models are still missing the 1M