OpenAI launches its new family of models with GPT-5.6

OpenAI didn't just drop new models. It fired a priced-to-move salvo across Anthropic's bow. The three-tier GPT-5.6 family — Sol, Terra, Luna — lands with benchmark receipts and a token budget that makes Anthropic's Fable lineup look expensive. Sol hits 80 on the Artificial Analysis Coding Agent Index. Fable 5 sits at 77.2. The gap is 2.8 points. The cost gap is wider: Sol runs at roughly one-third the output token spend, half the latency, half the tokens. Terra matches Fable 5. Luna beats Opus 4.8. OpenAI didn't leak these numbers. It published them.

The pricing table tells the real story. Sol at $5 input, $30 output per million tokens. Terra at $2.50 in, $15 out. Luna at $1 in, $6 out. Anthropic's comparable tiers sit higher. Enterprise buyers weigh total cost of ownership, not just benchmark peaks. OpenAI handed them a spreadsheet that favors the incumbent. Sam Altman's CNBC claim — 54 percent token efficiency gains on coding tasks — is a number designed for procurement officers, not researchers. It converts to dollars fast.

Cybersecurity is the sharp edge. The Trump administration sought to restrict this rollout. That detail alone signals capability. The company frames 5.6 as its strongest security model yet, citing frontier performance with fewer tokens. The use cases are defensive: threat modeling, code review, patching, blue teaming. Simulated attacks on your own infrastructure. That's a productized red team. It also happens to be a dual-use capability. The administration's intervention attempt failed. The model ships anyway. That's a precedent.

ChatGPT Work arrives same day. Desktop, web, mobile. Documents, spreadsheets, presentations. It's Copilot by another name, but baked into the ChatGPT surface that millions already touch. Distribution is the moat. Microsoft owns the OS layer. OpenAI owns the chat layer. Work turns that habit into an enterprise subscription. The timing isn't accidental. SpaceXAI and Meta dropped models this week. The market is flooding. OpenAI needs a workplace anchor that doesn't require an IT rollout cycle.

Anthropic built its reputation on enterprise trust. Constitutional AI. Safety-first messaging. Contracts with Fortune 500. OpenAI's response isn't philosophy. It's a benchmark index, a price list, and a cybersecurity credential. The Coding Agent Index becomes the scoreboard. OpenAI wrote the rules, ran the game, posted the win. Independent verification will follow. But the narrative is set: Anthropic's coding crown has a challenger that costs less and runs faster.

Token efficiency compounds. Half the output tokens means half the inference compute. Half the latency means developers stay in flow. One-third the cost means budget holders approve pilot programs without board meetings. These are compounding advantages. They don't show in a single benchmark run. They show in six-month ROI calculations.

The three-tier structure mirrors cloud compute: premium, standard, economy. Sol for production workloads. Terra for staging. Luna for experimentation. Developers can cascade down without switching APIs. That lock-in is deliberate. The API surface stays identical. Only the model identifier changes. Migration friction approaches zero.

OpenAI's moat has always been distribution. ChatGPT, Codex, API — three channels, one model family. Anthropic runs primarily through API and Claude interface. Microsoft bundles OpenAI into GitHub Copilot, Azure, Office. The distribution graph compounds daily. GPT-5.6 widens the pipe.

The cybersecurity restriction attempt reveals something about where power sits. When a sitting administration tries to slow a commercial model release over misuse potential, the model has crossed a threshold. Blue teaming automation at scale changes the offense-defense calculus. Defenders get the same tools. Attackers get them too. The net effect is unknown. But the genie is out. OpenAI shipped anyway. That decision — to ship despite political pressure — may matter more than the benchmark scores.

Enterprise buyers now face a cleaner choice. Anthropic sells trust and constitution. OpenAI sells throughput and price. The market will split. Regulated industries may stay with Anthropic's audit trail. Speed-obsessed shops will flip. The middle will run both, A/B testing in production. OpenAI's pricing makes that test cheap.

The week's competitor drops — SpaceXAI, Meta — add noise. They don't change the binary. The AI race has two credible enterprise vendors. Everyone else is fighting for third. GPT-5.6 cements that duopoly. The benchmark war is now a price war. OpenAI fired the opening salvo. Anthropic's move comes next.