Between June 30 and July 9, every price lane a founder routes to moved at once. Anthropic repriced the mid-tier, OpenAI took its new family live, xAI and Meta undercut both, and Google made generated media almost free. None of it was coordinated; all of it points the same direction. Here's the whole board, then the map for what to route where.
The board, on one screen#
| Model | Vendor | In / Out (per 1M) | The hook |
|---|---|---|---|
| Muse Spark 1.1 | Meta | $1.25 / $4.25 | self-managing 1M context; US preview |
| Grok 4.5 | xAI | $2 / $6 | $0.50 cached; native in Cursor; terse |
| Sonnet 5 | Anthropic | $2 / $10 (promo) | near-Opus; → $3 / $15 on Sep 1 |
| Terra (GPT-5.6) | OpenAI | $2.50 / $15 | balanced; 90% cache-read discount |
| Nano Banana 2 Lite | $0.034 / 1K images | ~4s per image | |
| Omni Flash | $0.10 / sec video | clips up to ~10s |
What each move actually was#
Anthropic — Claude Sonnet 5 (June 30). A near-Opus model at introductory $2 in / $10 out per 1M tokens through August 31, reverting to $3 / $15 after. TechCrunch framed it plainly: a cheaper way to run agents. The catch to diary: that $10 output rate is a promo. If Sonnet 5 carries real traffic, model your economics against Opus-vs-Sonnet at the post-August number, not the launch one.
OpenAI — GPT-5.6, GA July 9. The three-tier family — Sol (flagship), Terra (balanced, $2.50 / $15), Luna (cheap) — went generally available and became ChatGPT's default. The clean cheap→mid→frontier ladder inside one SDK is the point: where Terra sits against Sonnet 5 and Gemini Flash is now the live mid-tier question.
xAI — Grok 4.5 (July 8). $2 / $6 with a $0.50 cached-input rate, native in Cursor on every plan, and co-trained on agent telemetry to emit fewer tokens per step. Terseness compounds across a long run — which is why the sticker price doesn't decide it.
Meta — Muse Spark 1.1 (July 9). The cheapest output token on the board at $4.25 / 1M, plus a self-managing 1M-token context that deletes plumbing you'd otherwise build. The asterisks: US-only public preview, no SLA, and it trails the leaders on the hardest coding. The three-way against Terra and Grok is where the real routing call lives.
Google — the media tier (June 30). Not a chat model — a generation tier. Nano Banana 2 Lite makes an image for $0.034 per thousand and Omni Flash makes video for ten cents a second, both on the Gemini API. Cheap enough to generate per request instead of pre-rendering a library.
The routing map#
The mistake this week invites is picking the lowest sticker price. Don't. The board rewards a workload-first read:
- Rank agentic paths by output-tokens-per-task, not per-token price. Every model here prices output several times above input, and agents are output-heavy. A terser model at a higher rate (Grok) can beat a chatty one at a lower rate (Muse Spark) on cost-per-finished-job. Run your own eval, count output tokens, multiply.
- Cheap, high-volume, long-context agent work → Muse Spark 1.1, and let its self-managing context carry the state.
- In-editor coding loops → Grok 4.5, where terseness and Cursor distribution both land.
- General mid-tier production traffic → Sonnet 5 (watch the September reprice) or Terra (if you're on OpenAI tooling and cache-heavy).
- The hardest reasoning and coding → a frontier tier (Sol, Opus). None of the cheap four are built to be that, and their vendors don't claim they are.
- Any picture or short clip your product ships → generate it on demand on Google's media tier; at $0.000034 an image, a static asset is now the expensive option.
The week's real lesson isn't that a new cheapest model arrived. It's that "cheapest" stopped being a number on a price sheet and became a property of your specific workload — and the only way to read it is to run the tokens.
