fix(core): price Copilot long-context tiers and cache writes by golfromeo-fr · Pull Request #59803 · anomalyco/opencode

golfromeo-fr · 2026-10-28T02:17:02Z

Issue for this PR

Closes #52089
Closes #59671

Type of change

Bug fix
New feature
Refactor / code improvement
Documentation

What does this PR do?

GitHub Copilot costs were undercounted badly on long sessions. In my case (#59671): a 3h session on Copilot / Claude Haiku 5.5, 348 requests, ~344K average context, OpenCode showed $1.56 when the real bill is roughly $9.34 (between $2.03 and $10.15, since Copilot picks the price tier per request by prompt size).

Two causes, both in the Copilot provider code:

Copilot's /models payload ships two price tiers (default and long_context, the latter starting above default.max_prompt_tokens). The builder only read the default tier.
The builder hardcoded cache writes to $0 (cache_write_price is in the payload, unread), and both usage adapters dropped cache_write_tokens, so cache writes weren't counted at all.

Pricing that session's totals at the ≤100K rates with free writes gives $1.5557 — cent-exact with the displayed $1.56, which is how I know exactly which rates the old code applied.

The builder now emits the write price and a context-tier cost entry, which SessionUsage.calculateCost already applies per request, so no calculator change was needed. The adapters map cache_write_tokens into cacheWrite. While there I replaced their hand-rolled noCache arithmetic with ProviderShared.subtractTokens/sumTokens like the shared protocols use: three of the four paths left noCache undefined on responses without cache details, which downstream prices the entire prompt at $0, and none of them clamped against broken breakdowns.

Deliberately not addressed: Copilot's own reported billed amount (#35765) and the richer pricing shape from #59460.

How did you verify your code works?

New unit tests against the Copilot rate card for Claude Haiku 5.5: long-context tier applies above 100K prompt tokens, default tier at exactly 100K, cache writes priced, reasoning billed as output, a malformed long_context object no longer drops the model from the catalog, cache-write mapping in all four adapter paths, plus regressions for the $0-prompt and negative-noCache cases. The adapter regressions fail on the parent commit and pass with this change. bun run check passes, and the packages/core suite has the same 4 failures as a clean dev checkout (all environment-related).

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" role="img" aria-label="Copilot session cost: displayed $1.56, actual about $9.34">
  <style>
    .bar { animation: grow 2.8s ease-in-out infinite; }
    .tag { animation: fade 2.8s ease-in-out infinite; }
    @keyframes grow { 0%, 15% { width: 26px } 55%, 100% { width: 142px } }
    @keyframes fade { 0%, 30% { opacity: 0 } 65%, 100% { opacity: .9 } }
    @media (prefers-reduced-motion: reduce) { .bar, .tag { animation: none } }
  </style>
  <g fill="currentColor" font-family="ui-monospace,SFMono-Regular,monospace" font-size="11">
    <text class="tag" x="12" y="11" font-size="9" opacity="0">6x undercount: tier + cache writes</text>
    <text x="12" y="27" opacity=".5">billed</text>
    <rect x="62" y="18" width="26" height="9" rx="2" opacity=".5"/>
    <text x="94" y="27" opacity=".5">$1.56</text>
    <text x="12" y="53">actual</text>
    <rect class="bar" x="62" y="44" width="142" height="9" rx="2" opacity=".9"/>
    <text x="206" y="53">$9.34</text>
  </g>
</svg>

Screenshots / recordings

Not a UI change, so none attached.

Checklist

I have tested my changes locally
I have not included unrelated changes in this PR

Added by the opencode bot (AI-generated).

Fixes #51845

Related: #35765, #53521

Copilot's /models payload exposes per-request price tiers (default plus long_context with its own prices and max_prompt_tokens) and a cache_write_price, but the model builder read only the default tier and hardcoded cache writes to zero, and both usage adapters dropped cache_write_tokens entirely. Long sessions on tiered models like Claude Haiku 5.5 were undercounted ~6x. The builder now emits the cache write price and a context-tier cost entry bounded by the default tier's max_prompt_tokens, which SessionUsage.calculateCost already applies per request. Both Copilot adapters map cache_write_tokens (or Anthropic-style cache_created_input_tokens) into cacheWrite and keep the non-cached input consistent with the total.

github-actions · 2026-10-28T02:17:13Z

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

Open an issue describing the bug/feature (if one doesn't exist)
Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

github-actions · 2026-10-28T02:50:24Z

Thanks for updating your PR! It now meets our contributing guidelines. 👍

opencodejr

This looks good to merge. I ran the #52089 reproduction (a local mock of Copilot's /models and /responses) on the base and with this change:

Prices: cache writes go from $0 to $5/M, and a long-context tier is added above default.max_prompt_tokens.
Usage: the Responses generate and stream paths now report cacheWrite: 8270 and noCache: 3, where the base reported no cacheWrite and noCache: 8273.

The tier boundary matches how SessionUsage.calculateCost picks tiers (input plus cache reads plus cache writes, strictly above the boundary). The Copilot and session-usage tests pass. I couldn't run the package type check in this environment.

Public API is unchanged: only the internal parsing of Copilot's /models and usage responses changes.

One small point is inline: the new session-usage.test.ts cases exercise unchanged code and can be dropped.

Fixes #51845
Fixes #59671
Related: #35765, #53521

opencodejr · 2026-10-28T02:54:08Z

  })
 })
+
+// GitHub Copilot's Claude Haiku 5.5 rate card: a per-request tier above 100K prompt tokens at 5x.


These four cases test SessionUsage.calculateCost, which this PR doesn't change, and they pass on the base too. The new coverage is the tier and cache-write prices from CopilotModels, which models.test.ts already checks. Could you drop these so the change stays focused?

Done in 238a59b — dropped the session-usage.test.ts cases. The changed code stays covered by the builder and adapter tests.

Review feedback on anomalyco#59803: the tier and cache-write pricing behavior they pinned lives in SessionUsage.calculateCost, which this PR does not change, and the changed code is already covered by the builder and adapter tests.

opencodejr

The new commit removes the session-usage.test.ts cases I flagged, which addresses my earlier comment. It changes nothing else. I re-ran the Copilot tests and the #52089 and #51845 reproductions at this head: cache-write prices, the long-context tier and the cacheWrite/noCache mapping are all correct. No new concerns. The public API is unchanged.

opencodejr

The new commit only merges dev. I checked that the PR's own changes are the same as before and that the merged dev commits don't touch the Copilot provider or the cost calculation. I re-ran the Copilot tests and the #52089 and #51845 reproductions at this head, and all are still fixed. My earlier test comment was addressed in the previous push. No new concerns, and the public API is unchanged.

github-actions Bot added needs:compliance This means the issue will auto-close after 2 hours. needs:issue labels Oct 28, 2026

github-actions Bot removed needs:issue needs:compliance This means the issue will auto-close after 2 hours. labels Oct 28, 2026

golfromeo-fr marked this pull request as ready for review October 28, 2026 02:51

opencodejr Bot reviewed Oct 28, 2026

View reviewed changes

github-actions Bot mentioned this pull request Oct 28, 2026

📊 AI CLI 工具社区动态日报 2026-10-28 96loveslife/big_model_radar#716

Open

Merge branch 'dev' into copilot-tier-pricing

59a8099

opencodejr Bot reviewed Oct 29, 2026

View reviewed changes

Navigation Menu

Repository navigation

fix(core): price Copilot long-context tiers and cache writes - #59803

fix(core): price Copilot long-context tiers and cache writes#59803
golfromeo-fr wants to merge 3 commits into
anomalyco:devfrom
golfromeo-fr:copilot-tier-pricing

golfromeo-fr commented Oct 28, 2026 •

edited by opencodejr Bot

Loading

Uh oh!

github-actions Bot commented Oct 28, 2026

Uh oh!

github-actions Bot commented Oct 28, 2026

Uh oh!

opencodejr Bot left a comment

Uh oh!

opencodejr Bot Oct 28, 2026

Uh oh!

golfromeo-fr Oct 28, 2026

Uh oh!

opencodejr Bot left a comment

Uh oh!

opencodejr Bot left a comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

golfromeo-fr commented Oct 28, 2026 • edited by opencodejr Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Issue for this PR

Type of change

What does this PR do?

How did you verify your code works?

Visual

Screenshots / recordings

Checklist

Uh oh!

github-actions Bot commented Oct 28, 2026

Uh oh!

github-actions Bot commented Oct 28, 2026

Uh oh!

opencodejr Bot left a comment

Choose a reason for hiding this comment

Uh oh!

opencodejr Bot Oct 28, 2026

Choose a reason for hiding this comment

Uh oh!

golfromeo-fr Oct 28, 2026

Choose a reason for hiding this comment

Uh oh!

opencodejr Bot left a comment

Choose a reason for hiding this comment

Uh oh!

opencodejr Bot left a comment

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

golfromeo-fr commented Oct 28, 2026 •

edited by opencodejr Bot

Loading