What fails#
From the 2026-11-17 21:51 UTC poll, OpenCode reports ingest_failed: opencode: GPT Image 2.5 Sunburst has an endpoint but no price row (source https://raw.githubusercontent.com/anomalyco/opencode/dev/services/site/src/docs/content/console/models.mdx, PostHog project 643425). It repeats once per docs TTL.
Why#
The docs added image models to the endpoints table (line 121: gpt-image-2.5-sunburst → /inference/openai/v1/images/generations). Their prices are in a new ### Image pricing table, not the token price table that src/server/providers/opencode-pricing.ts reads:
| Model | Price | Notes |
|---|
| GPT Image 2.5 Sunburst | $1.00 / image | Plus $8 / $30 per 1M input / output tokens, and $2 per 1M cached input tokens. |
| Nano Banana 2.1 | $1.50 / $7.50 / $30.00 per 1M tokens | Input / text output / image output. About $0.034 per 1K image, $0.050 per 2K, $0.113 per 4K. |
| Grok Imagine Image 2.0 | $0.04 / image | $0.01 per input image; quality and resolution can raise the output price to $0.05–$0.08. |
The throw at opencode-pricing.ts:200 fails closed, which is correct.
Fix#
- Parse the
### Image pricing table into rate cards: per-image plus token levers for Sunburst; input, text-output and image-output token rates for Nano Banana 2.1; per-image plus per-input-image for Grok Imagine Image 2.0.
- Where a price is a range that depends on quality or resolution and the doc gives no rule (Grok's "$0.05–$0.08"), don't make up a value. Price the stated parts and mark the rest as
unverified or a range, as the rate-card rules say (no pricing fallbacks).
- Treat any unreadable image price row as a throw, the same as today.
- Tests from the three live rows.
Fixed by #690, deployed to prod as 9237af31 at 2026-11-17 22:46 UTC.