OpenAI has released the GPT-5.6 family in three tiers, Sol, Terra, and Luna, promising frontier performance with far fewer tokens. The pitch aims squarely at Anthropic, with OpenAI claiming flagship Sol beats Claude Fable 5 on coding while costing less. It also debuts ChatGPT Work, an enterprise companion app.

Key Takeaways

• GPT-5.6 launches in three tiers: Sol, Terra, and Luna
• OpenAI claims Sol beats Anthropic’s Fable 5 on coding
• Sol is billed as its strongest cybersecurity model yet
• Pricing spans $5/$30 (Sol) down to $1/$6 (Luna) per million tokens
• A new enterprise app, ChatGPT Work, launches alongside it

What OpenAI Announced

The release is a whole family, not a single model. OpenAI unveiled GPT-5.6 on Thursday in three variants: Sol, Terra, and Luna, pitching the set as a step up across enterprise work, coding, and scientific research.

The naming has a logic. OpenAI frames the number as the generation and the names as durable capability tiers that advance on their own cadence, so Sol, Terra, and Luna are meant to persist across future updates rather than being one-off releases.

The rollout followed a preview. A limited preview began on June 26 for approved partners, before the full global release went live on July 9 across ChatGPT, Codex, and the OpenAI API, rolling out gradually over the following day.

The timing was pointed. The launch landed the same week as OpenAI’s own GPT-Live voice models and rival releases from SpaceXAI and Meta, capping one of the busiest stretches of model launches yet.

The Three Tiers and Pricing

Each tier targets a different need. Sol is the workhorse flagship for the hardest tasks, Terra is the balanced middle option with GPT-5.5-level performance at roughly half the cost, and Luna is the fastest, most affordable choice for high-throughput routine work.

The pricing ladder is steep from top to bottom, quoted per million tokens:

• Sol: $5 input / $30 output
• Terra: $2.50 input / $15 output (exactly half of Sol)
• Luna: $1 input / $6 output

Efficiency is the headline selling point. CEO Sam Altman said the new models are far more cost-effective than before, telling CNBC that Sol is 54% more token efficient on AI coding tasks, and the family adds a 90% discount on cached input reads to trim costs further.

The Benchmark Claims (and the Anthropic Target)

The marketing is openly aimed at one rival. OpenAI calls Sol its best coding model yet and explicitly measures it against Anthropic’s much-hyped Fable 5.

The coding claim is specific. Using the Coding Agent Index, OpenAI says Sol sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about a third less. It adds that Terra performs just above Fable 5 and Luna outperforms Opus 4.8.

Other evals echo the pitch. OpenAI reports Sol scored 53.6 on Agents’ Last Exam, surpassing Fable 5 by 13.1 points, and set new marks on BrowseComp (92.2%) and OSWorld 2.0 (62.6%). On the independent Artificial Analysis Intelligence Index, though, Sol with max reasoning came within one point of Fable 5 rather than clearly beating it.

One important caveat runs through all of this: these are OpenAI’s own reported figures. They read as vendor evidence, and the closer Artificial Analysis result is a useful reminder to test the claims rather than take them as settled.

The Cybersecurity Angle

Security is where OpenAI leans hardest. It calls GPT-5.6 its strongest cybersecurity model yet, achieving frontier performance with significantly fewer tokens.

The gains are large on OpenAI’s benchmarks. The company says GPT-5.6 scored 73.5% on one exploit benchmark versus GPT-5.5’s 47.9% at a comparable token budget, and nearly doubled its predecessor’s pass rate on a separate timed exploit test.

The framing is defensive. OpenAI stresses the model supports threat modeling, code review and patching, and blue teaming, the practice of simulating attacks on your own systems to find weaknesses first.

This capability has political history. The Trump administration previously sought to restrict the model’s rollout over its cyber abilities, and OpenAI says GPT-5.6 is more capable in biology and cybersecurity than earlier models but does not cross its Critical risk threshold, backed by layered safeguards and roughly 700,000 GPU hours of automated red teaming before launch.

ChatGPT Work: The Enterprise Play

Alongside the models, OpenAI shipped a new app. ChatGPT Work is designed as a workplace companion for enterprise teams, running on desktop, web, and mobile.

Its job is everyday office work. The tool helps with daily clerical tasks like drafting documents, spreadsheets, and presentations, positioning ChatGPT as a hands-on assistant for routine business output rather than just a chatbot.

The strategy is clear. Anthropic has made itself the likable underdog by fixating on enterprise customers, and ChatGPT Work is OpenAI’s move to contest that ground directly, pairing capable models with a purpose-built business app.

The Competitive Context

The launch caps a frantic week in AI. GPT-5.6 followed closely on similar releases from SpaceXAI, whose Grok 4.5 also leaned on efficiency and price, and Meta, which is pushing its own chips and its Muse Spark models.

The battleground has shifted. Rather than chasing only raw capability, the leading labs are now competing hard on capability-per-dollar, with efficiency claims and tiered pricing doing much of the marketing work.

Anthropic remains the clear target. Nearly every GPT-5.6 comparison is drawn against Fable 5 or Opus 4.8, underscoring how much OpenAI wants to blunt its rival’s enterprise momentum.

Why It Matters

The real story is economics as much as intelligence. By splitting GPT-5.6 into three tiers and hammering token efficiency, OpenAI is betting that businesses care as much about cost and speed as about topping a benchmark chart.

The safety framing matters too. Shipping a model OpenAI calls its strongest for cybersecurity, while insisting it stays below a Critical risk threshold, shows the tightrope these labs now walk between capability and caution, especially after government scrutiny of exactly these abilities.

The takeaway for users is choice. If the tiers deliver as advertised, teams can match model to task, paying flagship prices only when they need Sol’s full power and dropping to Luna for routine work. Whether the benchmark claims hold up under independent testing is the question worth watching next.

Digital Trendings is your trusted source for AI news and updates, stay tuned for more.