Tech Stack
OpenAI shipped the GPT-5.6 family to ChatGPT, Codex, and the API on July 9, after a preview that started June 26. The models matter — but the naming system might matter more.
The naming change is the actual story
With GPT-5.6, the number identifies the generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence:
- -Sol — the flagship: advanced coding, scientific research, cybersecurity-grade work. $5 input / $30 output per million tokens.
- -Terra — the workhorse: performance comparable to the previous GPT-5.5 generation at roughly half the cost. $2.50 / $15.
- -Luna — the throughput tier for routine, high-volume tasks. $1 / $6.
If that structure looks familiar, it should — it's Haiku/Sonnet/Opus with different mythology. Every major provider has now converged on the same three-tier shape, because that's the shape production usage actually has: a little bit of hard, a lot of medium, and an ocean of easy.
The "tiers advance independently" part is the underrated piece. It means you can build against "Terra" as a stable capability contract and absorb generation upgrades without re-benchmarking your whole stack — the exact thing that makes routing tables maintainable.
The mid-2026 price map, in one place
Per million tokens (input/output), as of this week:
- -Luna: $1 / $6
- -Sonnet 5 (intro, to Aug 31): $2 / $10
- -Terra: $2.50 / $15
- -Sonnet 5 (from Sep 1): $3 / $15
- -Opus 4.8: $5 / $25
- -Sol: $5 / $30
- -Fable 5: $10 / $50
Read that list twice and the war is obvious: OpenAI priced Terra exactly at Sonnet 5's post-intro price, and Anthropic's intro window undercuts Terra by 20-33% for the next two months. The workhorse tier — where nearly all production agent spend lives — is now a knife fight.
What this means if you build on these APIs
First: cross-vendor routing is now worth the plumbing. When the workhorse tiers were priced apart, single-vendor simplicity won. At near-parity pricing, task-level benchmarking across two vendors pays for itself — I run my personal agents on a provider chain with automatic fallback for exactly this reason, and that pattern is about to become standard architecture rather than a power-user trick.
Second: capability tiers beat model names in your config. Structure your code so tasks declare a tier ("cheap", "workhorse", "flagship"), and map tiers to whatever model currently wins that slot. This week the workhorse slot might be Sonnet 5; in September it might be Terra. Your application code shouldn't care.
Third: Codex runs on GPT-5.6 now — if you're evaluating coding agents, last month's comparisons are stale. Every tool in the category (Claude Code, Codex, Cursor, and the rest) is riding a different model this month than last. Re-run your own evals; don't trust June's blog posts, including mine.
The uncomfortable truth for builders
Model costs at the workhorse tier have fallen roughly 50-80% in six months while capability went up. If your AI product's unit economics were marginal in January, they're healthy now — and if your pricing to clients hasn't moved, this price war is quietly widening your margin every quarter. The businesses that lose here are the ones still hard-coding one model name into everything and re-platforming in a panic every release day.
Naming finally grew up. Your architecture should too.
Want to build something like this?
I architect and deploy end-to-end AI systems — from MVP to revenue.
Let's Talk