OpenAI is pulling back the curtain on GPT-5.6, its latest model series, giving a small circle of trusted partners the first look before a wider release expected in the coming weeks. The new series splits into three distinct variants — Sol, Terra, and Luna — each targeting a different point on the capability-versus-cost curve. It’s the most structured tiering OpenAI has done yet, and the pricing tells an interesting story about where the AI market is heading.
- GPT-5.6 launches in three tiers — Sol, Terra, and Luna — each targeting a different balance of power and cost.
- OpenAI spent 700,000 GPU hours hardening GPT-5.6 against jailbreaks, making it the company’s most security-focused release yet.
- The US government received an early preview before today’s announcement, reflecting a new and controversial AI oversight arrangement.
- Sol is priced at $5 per million input tokens — half the cost of Anthropic’s now-suspended Fable 5 model.
Table of Contents
Meet Sol, Terra, and Luna: What Each GPT-5.6 Variant Actually Does
Sol sits at the top of the stack. OpenAI describes it as its strongest model to date, and it’s the one the company has clearly poured the most engineering effort into. It introduces a ‘max’ reasoning effort setting, which gives the model more time to think through complex problems before responding — a meaningful upgrade for tasks where accuracy matters more than speed. OpenAI also says Sol is its best model for cybersecurity work, specifically flagging its ability to help users identify and patch vulnerabilities. That’s a pointed capability to highlight, and it signals that OpenAI is seriously courting enterprise security teams as a customer base.
Terra slots in as the everyday workhorse. OpenAI claims its performance is comparable to GPT-5.5 — the previous generation — but at half the price. For companies that built workflows on GPT-5.5, Terra is essentially a cost-reduction opportunity with minimal migration friction. Luna rounds out the trio as the budget option, aimed squarely at high-volume, cost-sensitive use cases where raw power takes a back seat to economics.
On pricing, Sol comes in at $5 per million input tokens and $30 per million output tokens. Terra is $2.50 input and $15 output. Luna drops further to $1 input and $6 output. These numbers are worth putting in context: Anthropic’s Fable 5, before it was pulled from the market, cost $10 per million input tokens and $50 per million output. OpenAI is pricing Sol at literally half that rate for what it’s positioning as a comparable — or superior — capability tier. That’s not a subtle competitive move.
The Government Preview: Necessary Caution or a Worrying Precedent?
Before today’s announcement, OpenAI briefed the US government on GPT-5.6 and its capabilities. The limited partner preview that’s happening now was also done, OpenAI says, at the administration’s request — with participation shared back to the government. That’s a notable arrangement, and OpenAI seems at least mildly uncomfortable with it. The company stated plainly in its announcement: ‘We don’t believe this kind of government access process should become the long-term default.’
The context here matters. Earlier this month, President Trump signed an AI cybersecurity executive order asking companies to voluntarily submit their most powerful models for government review 30 days before public release. According to reporting from The New York Times, OpenAI, Anthropic, Google, xAI, and Microsoft had already been cooperating with this kind of pre-release access even before the order was formalised. The notable holdout was Meta, which the government has reportedly been pressuring to fall in line.
OpenAI is framing its cooperation as a pragmatic short-term concession — something it’s doing to clear the path for a rapid public release rather than out of any deep ideological alignment with government oversight of AI. Whether that framing holds up over time is another question. Once a process like this becomes routine, it tends to stick around. The AI industry should probably pay close attention to how this dynamic evolves over the next year.
GPT-5.6 and the Jailbreak Arms Race
One of the more striking things about this GPT-5.6 release is how much energy OpenAI put into security before launch. The company says it spent 700,000 GPU hours specifically hunting for universal jailbreaks — methods that could force the model to ignore its safety training across a wide range of inputs. That’s a serious investment of compute, and it suggests the threat landscape around frontier models has become genuinely difficult to manage.
OpenAI trained all three GPT-5.6 variants to reject what it calls ‘prohibited cyber assistance,’ which includes attempts to manipulate the model into bypassing its own guardrails. Beyond the pre-launch hardening, the company is committing to a rapid-response process to identify, assess, and fix newly discovered jailbreaks after the models go live. That’s an acknowledgment that no amount of pre-release testing catches everything — a realistic stance, but also one that puts ongoing pressure on OpenAI’s safety team.
The shadow of Anthropic’s recent crisis looms over all of this. A few weeks ago, Anthropic suspended access to its Mythos 5 and Fable 5 models entirely after the government stepped in. Reports indicated that Amazon and other companies had flagged to authorities that the models could be exploited for malicious purposes via jailbreaking. Anthropic has started restoring access — the US government has now given the company permission to redeploy Mythos to a select group of organisations — but the episode was a public embarrassment and a clear warning to everyone else in the space. OpenAI’s 700,000 GPU-hour jailbreak effort reads, at least in part, as a direct response to watching that unfold.
What This Means for the Broader AI Market
The structure of the GPT-5.6 launch reflects something broader that’s been building across the industry. Top-tier AI labs are no longer competing purely on capability benchmarks — they’re competing on price, safety posture, and government relationships simultaneously. OpenAI’s decision to build a three-tier model family rather than a single flagship isn’t just a product decision; it’s a market strategy designed to hold territory at every price point while Anthropic deals with regulatory turbulence and Google continues pushing Gemini into enterprise workflows.
The aggressive pricing on Sol compared to Fable is worth watching closely. If OpenAI can credibly claim comparable or better performance at half the cost, that’s a serious pitch to enterprises that are already running cost-benefit analyses on their AI spend. The broader public release, expected within weeks, will be the real test — that’s when independent benchmarks, developer feedback, and real-world performance will either validate or complicate the positioning OpenAI has put out today.
Source: Engadget
