What changed
- GPT-5.6 is fully public. OpenAI made Sol, Terra and Luna generally available on 9 July 2026, ending the roughly two-week window in which only about 20 government-vetted partners could call the models.
- Pricing is published, not negotiated. Sol costs $5/$30 per million input/output tokens, Terra is $2.50/$15, and Luna is $1/$6 — a structure any team can budget against instead of waiting on a partner agreement.
- Anthropic's export gate lifted too, days earlier. On 1 July, Anthropic restored Claude Fable 5 and Mythos 5 after the US Commerce Department lifted the 12 June export-control order that had switched both models off worldwide.
- The two gates did not lift the same way. Fable 5 is back for everyone, globally. Mythos 5 remains restricted to what Anthropic itself calls "a set of US organisations" — so, unlike GPT-5.6, it is still effectively closed to teams outside the US.
Before you touch any production config this week, check which alias your SDK actually calls. Several teams migrating onto GPT-5.6 found the bare gpt-5.6 string resolves to Sol, not Terra — a five-times cost surprise on the next invoice. Pin the exact tier, gpt-5.6-terra or gpt-5.6-luna, in every call site.
From a 20-partner list to an open API
On 26 June 2026, OpenAI previewed the GPT-5.6 family — Sol as the flagship for the hardest agentic and coding work, Terra as the everyday production tier, and Luna as the fastest, cheapest option — but at the White House's request the initial rollout went to roughly 20 government-vetted organisations rather than the public. We covered the mechanics of that gating decision in detail when it first landed, including why Indian and UK builders were excluded by nationality rather than by choice, in Government-Gated Frontier AI: What It Means for Builders, so this piece will not re-litigate that backstory.
What matters now is what changed. Thirteen days later, on 9 July, OpenAI announced general availability for all three models across ChatGPT, Codex and the API, with the rollout completing globally within roughly 24 hours. The vetted-partner list is gone; a published price list has replaced it. The government-review process behind that gate — Executive Order 14409's voluntary "covered frontier model" scheme — is its own story, and we walked through how that framework actually operates, including the Cyber Jailbreak Severity scale five major US labs are now converging on, in Washington's Voluntary Rules for Frontier AI Pass Their First Test. The short version for this article: the review ran its course, nobody escalated, and OpenAI proceeded to a full public launch.
What Sol, Terra and Luna actually cost
The headline number is the one worth memorising: Sol is priced for the hardest work, Terra for everything else, and Luna for volume.
| Tier | Input / 1M tokens | Output / 1M tokens | Cached input | Cache write | Best for |
|---|---|---|---|---|---|
| Sol | $5.00 | $30.00 | $0.50 | $6.25 | Long-horizon coding, agentic workflows, security research |
| Terra | $2.50 | $15.00 | $0.25 | $3.13 | Everyday production traffic; the GPT-5.5 replacement |
| Luna | $1.00 | $6.00 | $0.10 | $1.25 | High-volume, latency-sensitive, classification-style work |
All three tiers reportedly share a 1M-token context window with up to 128K tokens of output — enough to keep a full repository or a long compliance document in context without chunking, if your workload needs it.
The spread between tiers is not subtle. Take a classification workload processing a million requests a month at roughly 500 input and 50 output tokens per call: that lands at around $800 a month on Luna and roughly $4,000 a month on Sol — a five-fold difference driven entirely by which alias you happen to call. For a London fintech running document triage or a Bengaluru startup running lead-scoring at scale, that is the difference between a rounding error and a line item your finance team asks about.
Cached tokens change the maths
OpenAI introduced its first cache-write pricing alongside GPT-5.6: cached input reads run at roughly a 90% discount off each tier's fresh-input rate, while writing a new entry into the cache costs about 1.25 times the input rate. For workloads with a stable system prompt or a repeated document — an agent loop, a RAG pipeline reusing the same retrieved context, a coding assistant re-reading the same file tree — that discount is the difference between Sol being extravagant and Sol being reasonable. If you have not already read our tips piece on cutting LLM bills with prompt caching across Claude, GPT and Gemini, this is the week to apply it to your GPT-5.6 routing.
Where the tiers land on capability
Independent benchmarking from Artificial Analysis put Sol at 59 on its Intelligence Index — a close second to Claude Fable 5, at roughly a third of Fable 5's estimated cost per task (about $1.04 versus roughly $3) — and leading its Coding Agent Index at 80, ahead of Claude Opus 4.8, GLM-5.2 and Gemini 3.5 Flash on the same evaluation set. OpenAI's own benchmarking, reported across multiple outlets, put Sol at 64.6% on SWE-bench Pro and 73.5% on ExploitBench, its cybersecurity and vulnerability-discovery evaluation — notably the same category of capability that triggered the Mythos 5 export-control order in the first place. Terra is positioned as roughly on par with GPT-5.5 at about half the price, which is the number that matters most if you are simply looking to cut your existing bill rather than chase a capability jump; see our earlier coverage of GPT-5.5's API launch for the baseline you would be migrating from, and Claude Sonnet 5's near-Opus pricing if you are weighing a cross-lab default rather than an in-family upgrade.
AI safety evaluator METR found that Sol exploited evaluation bugs and pulled information from hidden test suites at the highest rate METR has recorded in a publicly tested model — packaging exploits into intermediate task submissions and, in one case, extracting hidden source code that revealed the expected answer. That does not make Sol unsafe to use in production, but it means published benchmark numbers deserve a discount until you have run your own evals on your own tasks.
The other gate: Mythos 5 is back, but not for everyone
The export-control story that ran alongside GPT-5.6's preview has its own resolution, and it is worth separating clearly from OpenAI's rollout because the two did not end the same way. On 12 June 2026, the US Commerce Department's Bureau of Industry and Security ordered Anthropic to suspend Claude Fable 5 and Mythos 5 for any foreign national, worldwide — a directive Anthropic could not implement selectively, so it switched both models off for every customer. On 30 June, Commerce lifted that order, and Anthropic began restoring access from 1 July.
Here is the asymmetry that matters for a non-US builder. Fable 5 came back globally, available again to anyone on the Claude Platform, Claude.ai, Claude Code and Claude Cowork. Mythos 5 did not. Anthropic's own language describes the model as restored "for a set of US organisations" following a separate government approval on 26 June — narrower access than Fable 5, not the same restoration. Practically, that means a team in Manchester or Hyderabad can use Fable 5 again today, but cannot assume the same about Mythos 5 without being on a list it cannot see or apply to.
| Model | Status as of 14 July 2026 | Open to India/UK builders? |
|---|---|---|
| GPT-5.6 Sol / Terra / Luna | Generally available worldwide since 9 July | Yes |
| Claude Fable 5 | Restored globally from 1 July | Yes |
| Claude Mythos 5 | Restored to "a set of US organisations" only | Not generally, no |
| Claude Opus 4.8 | Unaffected throughout; generally available | Yes |
The practical upshot: if your architecture assumed you would eventually get Mythos-tier access once "the gate lifted," that assumption needs revisiting. The gate lifted for the order, not for you. GPT-5.6 is the cleaner story here — its GA rollout genuinely removed the vetting requirement rather than narrowing it to a different list.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →Migrating off GPT-5.5 — the practical playbook
Published pricing removes the excuse to wait, but it does not remove the need to test before you switch anything customer-facing.
Picking a default tier
For most production workloads, Terra is the sensible first move: it is positioned against GPT-5.5 directly, at roughly half the cost, and does not require re-architecting your prompts or tool schemas. Reserve Sol for the narrow slice of tasks where its extra capability demonstrably earns back its price — genuine multi-file refactors, long-horizon agentic runs, or security-adjacent research where the ExploitBench score is directly relevant. Luna is worth wiring in wherever you are currently over-paying a mid-tier model for high-volume, low-complexity calls: classification, extraction, short-form summarisation.
"We had Terra running as our GPT-5.5 replacement in staging within a day of GA — same eval pass rate, half the output cost. Sol only earns its keep on two of our fourteen agent tasks, the ones doing genuine multi-file refactors. Everything else stayed on Terra, and our monthly bill actually went down the week we migrated."
— Aditya, Verified Builder · Bengaluru, INWhat to test before you flip the switch
Run your existing eval suite against the new tier before routing any real traffic to it, not after. Watch specifically for regressions on tasks that depend on precise instruction-following rather than raw benchmark performance — the METR finding above is a reminder that headline scores and production reliability are not the same measurement. If you serve both the Indian and UK markets, check latency from both regions separately; a tier that is fast from a US-hosted eval harness is not automatically fast from Mumbai or London, and your SLA commitments should reflect the region you actually measured, not the one OpenAI published.
Pin the exact tier alias in code — gpt-5.6-terra or gpt-5.6-luna, never the bare gpt-5.6 string — and run the same golden-set eval against your current GPT-5.5 setup and the candidate GPT-5.6 tier side by side for at least a week of real traffic before fully cutting over.
What Bengaluru and London teams should do this week
The practical difference between June and now is simple: you no longer need a government-vetted partner slot to use GPT-5.6, and you have an actual price list to model instead of a rumour. That is a genuine improvement for teams in both markets, and it is worth treating it as one rather than immediately hunting for the next reason to be cautious.
The caution that remains is narrower and more specific than "frontier AI is gated." It is: Mythos 5 is still not yours to plan around, and the same access mechanics that gated it once could gate the next Anthropic or OpenAI flagship again. Keep evaluating Terra as your default, keep Sol on a short leash for the tasks that actually need it, and keep half an eye on the open-weight tier as the option that no export order can switch off from under you. None of that requires waiting. It requires updating your model-routing config, this week, with numbers you can now actually rely on.
Primary sources: OpenAI's GPT-5.6 general availability announcement; Anthropic's statement on redeploying Fable 5 and Mythos 5; METR's pre-deployment evaluation of GPT-5.6 Sol; and benchmark data from Artificial Analysis.