Jul, 144 min to read
AI Products
News · OpenAI ships GPT-5.6 as three tiers priced for token efficiency
Jul, 94 min to read
AI ProductsOpenAI ships GPT-5.6 as three tiers priced for token efficiency
Sol, Terra, and Luna arrive with a naming scheme that separates generation from capability, benchmarks that lead on cost-per-result more than raw intelligence, and a cyber-access regime that requires hardware passkeys by September 1.
Three durable tiers, and a naming split worth noting
OpenAI released GPT-5.6 in three tiers: Sol as the flagship, Terra as a balanced everyday model, and Luna as the cheapest and fastest. API pricing per million tokens is $5/$30 for Sol, $2.50/$15 for Terra, and $1/$6 for Luna.
The more interesting structural decision is in the names. OpenAI states that the number identifies the generation while Sol, Terra, and Luna are 'durable capability tiers that can advance on their own cadence.' That decouples the marketing generation from the individual model's release schedule — a signal that OpenAI intends to update tiers independently rather than shipping a monolithic version bump each time.
Access is segmented across products too. Free and Go users get Terra in ChatGPT Work and Codex; Plus and above can pick among all three and set an effort level. The highest settings — max and the multi-agent 'ultra' mode — are gated to higher-paying plans.
The efficiency claims are stronger than the intelligence claims
Almost every headline result is framed as cost-per-result rather than pure capability. On the Artificial Analysis Intelligence Index, OpenAI concedes Sol with max reasoning 'comes within one point of Fable 5' — the table shows Sol at 58.9 versus Claude Fable 5 at 59.9 — while arguing the win is completing tasks in 61% less time at roughly half the estimated cost.
The benchmark tables reward a close read. On SWE-Bench Pro, Sol scores 64.6% while Claude Fable 5 and Claude Mythos 5 sit at 80% and 80.3% and Opus 4.8 at 69.2%. On GDPval-AA v2, Sol's 1,747.8 Elo trails Fable 5's 1,759.6. OpenAI is not claiming a clean sweep on raw quality; it is claiming it reaches competitive quality with fewer tokens, less latency, and lower spend.
Where the family does post clear gains is efficiency across tiers. On the Coding Agent Index, Sol sets 80 versus Fable 5's 77.2 'while using less than half the output tokens.' The customer quotes reinforce the same theme: Qodo reports roughly 3x fewer tokens per PR, Lovable cites 35–48% fewer tool calls, Model ML 39% fewer tokens per deck. The story OpenAI is selling is unit economics, not a decisive intelligence lead.
Orchestration moved inside the model
Two capabilities push agent coordination into the model and the API rather than into developer glue code. Programmatic Tool Calling lets GPT-5.6 write and run in-memory programs that coordinate tools, filter intermediate data, and choose the next action — reducing model round trips. OpenAI notes this path is Zero Data Retention compatible in the Responses API.
The 'ultra' setting coordinates four agents in parallel by default, with 16-agent configurations shown on BrowseComp and SEC-Bench Pro. OpenAI's framing is that parallel agents 'shift the score-latency frontier upward and to the left' — trading higher token spend for both stronger results and faster completion. Developers get this through the multi-agent beta in the Responses API.
Rogo's quote quantifies the tradeoff cleanly: with Programmatic Tool Calling it 'matched quality while using 24% fewer output tokens and completing tasks 28% faster.' For teams building production agents, the pitch is that behaviors they previously scripted are now first-class API primitives.
Cyber access gated behind hardware passkeys and a reasoning monitor
GPT-5.6's cybersecurity gains are large — ExploitBench 2 jumps to 73.5% from GPT-5.5's 47.9%, and ExploitGym roughly doubles peak pass rate — and OpenAI has paired them with a notably more restrictive access model. Sol's cyber safeguards 'block roughly ten times more potentially harmful activity' than prior models, with a reasoning monitor layered on top of trained-in protections and classifiers.
To keep working with the most cyber-capable settings, individuals in the Trusted Access for Cyber program must enable Advanced Account Security with hardware-backed passkeys by September 1 or drop to default access. OpenAI even arranged preferred Yubico pricing for users without keys. This ties frontier capability to a verified-identity hardware requirement — a concrete access-control decision, not just a policy statement.
OpenAI also discloses that GPT-5.6 does not cross its Critical threshold in biology or cybersecurity, and argues against overblocking because defenders and open-source models compete in the same space. The specific implication for teams: capability is increasingly bundled with account-level enforcement and verification. Planning a GPT-5.6 deployment now means budgeting for identity verification and the friction of retry-on-lower-model fallbacks, not only for token spend.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.