GPT-5.6 Arrives in Three Tiers: What Sol, Terra, and Luna Actually Change for Your Coding Stack

OpenAI shipped GPT-5.6 on July 9, and for the first time it published a genuinely granular price ladder: three tiers, spanning $1 to $30 per million tokens, each aimed at a different kind of buyer (TechCrunch). If you picked your coding model earlier this year and haven't looked since, the ground has shifted under you.

The release almost didn't happen on schedule. CNBC reported that the Trump administration had sought to restrict the rollout over cybersecurity concerns before OpenAI released the model with capabilities focused on defensive work such as threat modeling and code review (CNBC). That context matters less for your tooling decision than the numbers do, so let's get to those.

The three tiers, priced

Prices verified against OpenAI's API pricing page, which also lists batch pricing at half these rates - worth knowing if your workload tolerates async processing. All three tiers are live across ChatGPT, Codex, and the API (TechCrunch).

The headline benchmark claim: Sol scores 80 on the Artificial Analysis Coding Agent Index - 2.8 points above Anthropic's Fable 5 - while using under half the output tokens and, per Sam Altman, running "54% more token efficient" on coding tasks. Treat those as vendor claims until independent replications land, because launch-week benchmark numbers have a long history of shrinking under third-party measurement. But even discounted, the shape of the offer is clear: OpenAI is competing on efficiency-per-dollar, not just raw capability.

What this actually changes if you're choosing a model

Here's the part most launch coverage skips. Whether GPT-5.6 matters to you depends almost entirely on which of three buckets you're in.

You pay per token through the API. This is where the reset is real. A month ago, the budget-capable conversation was mostly Claude Sonnet versus older GPT tiers. Now three models cluster in the $1-6 per million output band: Luna ($1/$6), Claude Sonnet 5 at its intro price ($2/$10 through August 31, then $3/$15 - see our Claude pricing page), and Grok 4.5 ($2/$6). If your agent workload burns output tokens - and agentic coding workloads do - Luna's $6 output rate is now the number everyone else has to justify themselves against. Run a week of your real traffic through it before you commit; efficiency claims measured on benchmarks don't always survive contact with your codebase.

To make that concrete, take an agent pipeline pushing 100M output tokens a month - a mid-size team running agents daily. Output alone: Luna costs $600/month, Terra $1,500, Sol $3,000, Sonnet 5 $1,000 at intro pricing and $1,500 from September. On the year, the Luna-versus-Terra gap is $10,800 - real procurement money riding on whether the cheap tier holds up on your tasks. That's the experiment worth running this month, while Sonnet 5's promo window still muddies the comparison.

You pay for a coding assistant seat. Cursor, GitHub Copilot, and friends abstract the model behind a subscription, so the direct price change doesn't hit your invoice - but the model menu inside those tools will shift over the coming weeks, and with Copilot now metering usage through AI Credits, which model your agent selects quietly determines how fast your credits drain. If your tool lets you pin a model, it's worth re-checking whether your pin still makes sense. Our Claude Code vs Cursor vs Copilot comparison covers the seat-level decision.

You're on ChatGPT Plus and just want the best default. You'll get GPT-5.6 without doing anything. The tier routing happens on OpenAI's side. No action needed - this launch costs you nothing and asks nothing of you.

The comparison nobody should skip: Terra vs Sonnet 5

Anthropic made Claude Sonnet 5 its default model on June 30 with intro pricing of $2/$10 per million tokens - deliberately undercutting what OpenAI charged at the time. Nine days later, Terra landed at $2.50/$15. That's not a coincidence; it's a price war conducted in public, and the beneficiary is anyone spending real money on tokens.

The catch worth flagging: Sonnet 5's intro pricing expires on August 31, 2026, stepping up to $3/$15. After that date, Terra and Sonnet 5 cost nearly the same, and the decision reverts to capability fit rather than price. If you're building a cost model right now, build it on the September numbers, not the July ones - otherwise you'll be re-doing this exercise in six weeks.

What to do this week

• If you run agentic coding through an API, benchmark Luna against your current model on your own tasks. At $1/$6 the trial costs almost nothing.
• If you're on Sonnet 5's intro pricing, put August 31 in your calendar. Your effective cost rises 50% on output tokens that day.
• If you use Cursor or Copilot, watch for GPT-5.6 appearing in the model picker and re-check any pinned model choices.
• Ignore the launch-week benchmark discourse until Artificial Analysis and independent evals publish their own numbers on the shipped model. Vendor-reported deltas of 2.8 points are inside the noise band of methodology choices.

The honest bottom line: GPT-5.6 doesn't obviously dethrone anything yet - the claims need independent confirmation - but it repriced the market, and repricing is the part that affects your bill regardless of which benchmark wins. Check the current numbers on our ChatGPT pricing page before you decide.