[VickyXAI]

What changes

Chat — src/utils/constants.ts

• CHAT_PRICE_PER_MTOKEN: adds the 33 live, available, paid chat ids that had no row. Without a row an explicit call reserved the $5/$30 default, which shuts small budgets out of cheap models. Added: openai/gpt-6.1-sol, gpt-5.6-sol-pro, gpt-5.6-terra-pro, gpt-5.6-luna-pro, chat-latest, gpt-5.4, gpt-5.4-mini, gpt-5.2, gpt-4.1{,-mini,-nano}, gpt-4o{,-mini}, o3, o3-mini, o4-mini, google/gemini-3.8-flash, gemini-3.6-flash, gemini-2.5-pro, gemini-3.1-flash-lite, gemini-2.5-flash-lite, deepseek/deepseek-v4-flash-vision-exp, zai/glm-5.3, glm-5.3-flash, minimax/minimax-m2.7, qwen/qwen3.8-max, qwen3.8-flash, qwen3.7-plus, tencent/hy4-preview, xiaomi/mimo-v2.5, mimo-v2.5-pro, poolside/laguna-s-2.1, mistral/mistral-large-4.
• Removes tencent/hy3. It's hidden and the wallet gateways redirect it to qwen3.7-flash. Its $0.132/$0.528 row was stale against the account sheet ($0.0825/$0.33). Without a row it reserves the $5/$30 default, which covers every rail, including the redirect target's own long-context steps.
• Keeps anthropic/claude-opus-4 on purpose. It's hidden but the account rail still serves it at $15/$75, so removing the row would under-reserve.
• LONG_CONTEXT_PRICE_PER_MTOKEN: adds gpt-6.1-sol >272K → $4/$15, the only new step in the catalogue.
• MODEL_TIERS.free: re-probed both gateways (unpaid, ~3,000-char prompt, max_tokens 32, response model checked). cohere/north-mini-code and poolside/laguna-xs-2.1 now answer as nvidia/nemotron-3-super-120b on both chains, so they leave the tier. nvidia/nemotron-3-super-120b serves itself on both chains and joins it. New order: gpt-oss-20b, nemotron-3-super-120b, llama-3.2-11b-vision, nemotron-3-nano-omni (tail).
• FREE_CHAT_MODELS keeps gpt-oss-120b, step-3.7-flash, mistral-nemotron, nemotron-nano-* and gemma-4-31b. All fourteen $0 ids still answer HTTP 200 with no payment on both chains. The gateway lists them as billingMode:"free", and MODEL_REDIRECTS sends them to $0 targets. The account sheet also bills them at $0. If they were removed, an explicit call would reserve $5/$30 for a free call, and an exhausted budget would refuse it.

Image — src/tools/image.ts

• Adds openai/gpt-image-2.5-flare and -sunburst at $0.28 (1024²) and $0.56 (1536×1024 / 1024×1536). Sunburst joins the edit and mask sets, as in the gateway's image-editing.ts.
• Adds xai/grok-imagine-image-2.0 at $0.04.
• xai/grok-imagine-image-pro: $0.07 → $0.05. The gateway corrected this on 2026-09-08, so we had been over-reserving by 40%.

Video — src/tools/video.ts

• Adds xai/grok-imagine-video-1.5: $0.08 / $0.14 / $0.25 per second at 480p / 720p / 1080p, 1–15s, 8s default. No margin, same as grok-imagine-video. 4K is refused before payment.

Sources

• GET https://blockrun.ai/api/v1/models and https://sol.blockrun.ai/api/v1/models (2026-10-20). Both quote the same $/M input/output for all 88 chat models.
• Gateway registry blockrun/src/lib/models.ts: image sizes[] base prices, video pricePerSecondByResolution, free/hidden flags. src/lib/image-editing.ts for edit/mask sets. Chat route MODEL_REDIRECTS.
• Live unpaid 402 quotes from npm run verify:prices. They are pinned in the new tests.

Checks

• npm run typecheck: clean
• npm test: 1422/1422 pass
• npm run build: ok
• npm run verify:prices: exit 0. 0 under-reserving on Base and Solana. The catalogue sweep finds 0 gaps on Base, Solana and the account sheet. New probes: grok-imagine-video-1.5 matches exactly at every tier. The new image rows over-reserve only by the usual $0.001 fee gap.
• The schema-token test failed because the tool descriptions grew. README, docs/mcp-schema-overhead.md and the context-cost cards were regenerated: full profile 14,018 → 14,335 tokens, media 6,827 → 7,144, trading saving 61% → 62%.

package.json is untouched. No version bump, no publish.