OpenAI's GPT-5.6 Pricing Mirrors Claude's, Tier for Tier

Two three-tier ladders, nine days apart

Claude Sonnet 5 shipped June 30 with a promo price of $2 per million input tokens and $10 per million output, reverting to $3/$15 on August 31, sitting under Opus 4.8's $5/$25. We wrote about that ladder last week. Nine days later, on July 9, OpenAI shipped GPT-5.6 as three named models — Sol, Terra, Luna — with a price ladder that lands in almost the same place, tier for tier.

That's not a coincidence worth over-reading, and we won't pretend to know which lab priced against which. What's useful is that the two ladders now let you compare directly, model class for model class, instead of squinting at capability benchmarks and guessing at cost.

The tiers, side by side

Sol, the flagship, is priced at $5 input / $30 output per million tokens — same input price as Opus 4.8, but 20% pricier on output. Terra, the mid-tier, comes in at $2.50/$15 — cheaper on input than Sonnet 5's post-August list price ($3/$15), identical on output, but more expensive than Sonnet 5's current promo rate on both ends. Luna, the fast/cheap tier, is $1/$6; we don't have a clean Anthropic equivalent to line it up against, so we'll leave that comparison alone rather than force one.

The pattern that's worth naming: a three-tier structure — flagship, capable-and-cheaper, fast-and-cheapest — now looks like the shape both labs have converged on, at least for this generation. That's an interpretation, not a fact each company has confirmed, but the pricing tables from both sides line up closely enough that treating "which of three tiers" as the routing question, rather than "which vendor," is a reasonable default going into Q3.

Put a number on it: a workflow that burns 2M input and 500K output tokens a day costs roughly $25/day on Sol, $22.50/day on Opus 4.8, $12.50/day on Terra, and $13.50/day on Sonnet 5 at list price — or just $9/day while Sonnet 5's promo lasts. That promo figure is the one to watch, because it's temporary. Run the same math again on September 1 and Terra edges out Sonnet 5's list price by a dollar a day at this token mix, cheaper on input and exactly even on output. Neither vendor is uniformly cheaper once you factor in the promo window; the answer changes depending on which week you run the comparison, which is itself a reason to model cost against list price, not launch price, when you're picking a default for anything that ships past August.

GPT-5.6's system card also pairs the launch with what OpenAI describes as its most robust cybersecurity and biology safeguards yet, plus tighter access controls for high-risk use cases. We haven't tested how those classifiers behave under legitimate but adjacent workloads — security research tooling, bio-adjacent data pipelines — so we won't speculate on false-positive rates here. If your use case sits anywhere near those categories, budget time to test against the guardrails before you commit a routing path to Sol, not after.

The caching mechanics converged too

The more concrete news for anyone doing cost engineering: GPT-5.6 introduces explicit prompt-cache breakpoints, with cache writes billed at 1.25x the uncached input rate and cache reads getting a 90% discount — mechanically almost identical to what Claude's cache_control: ephemeral has done for over a year (4 breakpoints, 1.25x write multiplier, 90% read discount). We've written before about exactly this trade-off on Claude, and the math carries over almost unchanged to GPT-5.6.

One difference worth flagging for anyone routing traffic across both: GPT-5.6's cache life has a 30-minute minimum, versus Claude's 5-minute default TTL (extendable to 1 hour on request). For bursty agent traffic — requests clustered tightly, then idle — Claude's shorter default expires faster but costs less to opt into; GPT-5.6's longer floor means a cache you write once survives idle gaps you'd have lost on Claude's default. Neither is strictly better; it depends on how bursty your traffic actually is; worth a real trace, not a guess.

Concurrent subagents in a single request

The capability-side headline is that GPT-5.6 can run concurrent subagents and synthesize their output inside one API call — orchestration that currently lives outside the model on Claude, handled by the caller or a framework like the Agent SDK. We haven't run a side-by-side eval yet, so we won't claim one approach is faster or cheaper in practice. What we will say: if that pattern holds up under load, it's a meaningfully different default than "the caller owns the fan-out," and it's worth testing before you assume your existing orchestration layer is the only shape this can take.

What this means for your routing logic — again

Last week's article argued that Sonnet 5's effort parameter narrows the question from "which tier" to "how much effort does this specific call need." GPT-5.6 doesn't change that logic inside Claude. What it adds is a second axis on top of it: cross-vendor tier routing, where "cheap but capable" now has a real second candidate at a comparable price point, not just a fallback if your primary vendor has an outage. If your stack already does multi-vendor routing for redundancy, Terra is now a plausible default branch, not just a backup path — worth benchmarking against Sonnet 5's post-promo pricing specifically, since the promo window closes August 31.

What we're betting on next week

We're running our existing agentic-coding eval suite against Sol and Terra to see where the mirrored pricing actually holds up under real task difficulty, not just list price. We're also watching whether Anthropic moves on pricing before the Sonnet 5 promo expires August 31 — the timing makes that date worth watching for more than just a promo expiration.

→ Get in touch — we'll map your top use cases to the cheapest tier, across vendors, that still clears your quality bar.