26 июня 2026 — OpenAI ship'ит GPT-5.6 в трёх SKU: flagship Sol, balanced Terra, lightweight Luna (первый solar-system naming scheme). Tech deep dive для AI devs и tech leads: pricing matrix, Max/Ultra inference modes, TerminalBench 2.1 — 91,9% (#1 globally), CTF hit rate 96,7%, Cerebras 750 tok/s с июля, US government lock на preview, head-to-head vs Claude Mythos 5, 6-step access playbook и FAQ. Сейчас доступ только у ~20 pre-cleared partners; public rollout — через несколько недель.
Июнь 2026 — super release month на бумаге. На практике три blocker'а держат модель за paywall/government gate:
Access gated: по требованию US gov GPT-5.6 доступен только ~20 trusted partners через API/Codex — ни ChatGPT, ни public API для regular users
Competitive vacuum: Claude Mythos 5 offline с 12 июня (export control), Gemini 3.5 Pro slipped to July — coding agent market без clear SOTA
Policy entropy: executive order от 2 июня 2026 = precedent gov intervention в model releases; timeline prediction стал harder
| Model | Tier | Input | Output | Key metric |
|---|---|---|---|---|
| GPT-5.6 Sol | Flagship | $5 / M tok | $30 / M tok | TerminalBench 2.1: 91,9% |
| GPT-5.6 Terra | Workhorse | $2.50 / M tok | $15 / M tok | ~GPT-5.5 perf, −50% cost |
| GPT-5.6 Luna | Lightweight | $1 / M tok | $6 / M tok | 80% price edge vs Sol tier |
Current state: preview-only для ~20 partners. Polymarket odds на full release до 31 июля 2026: 87%.
Ночью 27 июня 2026 (CST) OpenAI анонсировала GPT-5.6 с solar naming: Sol (Sun), Terra (Earth), Luna (Moon) — flagship / mid / edge tier.
Release прошёл под government constraint: впервые US gov потребовал limited release до wide rollout. CEO Sam Altman compliant, но публично pushback:
Government approval gate не должен стать industry default — он отрезает лучшие tools от users, devs, enterprises и global partners, которым они нужны больше всего.
Strongest OpenAI model to date. Target workloads: hardcore coding, long-chain cyber research, multi-step agentic pipelines.
Два новых inference mode:
Pricing: $5/M input, $30/M output (parity с GPT-5.5).
Mass-deploy tier: support bots, internal tools, doc analysis. ~GPT-5.5 quality при 50% lower bill. $2.50/M in, $15/M out.
Tuned для high-QPS, low-latency: summarization, drafting, cron automation. Первый non-flagship с simultaneous High rating в cyber + bio. $1/M in, $6/M out.
| Model | Primary workload | Context | Cyber rating |
|---|---|---|---|
| Sol | Complex coding, security R&D, long agents | ~1.5M tok | High |
| Terra | Enterprise docs, support, mass API | ~1.5M tok | High |
| Luna | Summary, drafting, automation | ~1.5M tok | High |
89 hard CLI planning tasks; measures multi-step tool use, iterative fix loops, task coordination — closest proxy к real agent behavior.
| Model | Score | Mode |
|---|---|---|
| GPT-5.6 Sol | 91,9% — global #1 | Ultra (multi-agent) |
| GPT-5.6 Sol | 88,8% | Standard |
| Claude Mythos 5 | 88,0% | Standard |
| GPT-5.5 | 83,4% | Standard |
| Gemini 3.1 Pro Preview | 70,7% | Standard |
Sol dethroned Mythos 5 за 17 дней (Mythos topped chart 9 июня). Pre-release context: GPT-5.6 leak roundup.
| Model | Task completion (code mode) |
|---|---|
| GPT-5.6 Sol | 50,9% — единственный >50% |
| GPT-5.6 Luna | Slightly above GPT-5.5 |
Первая OpenAI product line, где все три SKU trigger High cyber risk tier.
| Model | CTF hit rate |
|---|---|
| Sol | 96,7% |
| Terra | 91,84% |
| Luna | 85,19% |
ExploitBench: Sol ≈ Anthropic Mythos Preview, но ~⅓ output tokens — materially cheaper security research stack.
Safety boundary: Sol находит vulns/primitives в Chromium/Firefox codebases, но не собирает full weaponized exploit chains autonomously — below OpenAI «Cyber Critical» threshold.
GPT-5.6 Sol на Cerebras WSE: до 750 tok/s generation. Baseline flagship: 50–150 tok/s. Math: 5–15× throughput — game changer для realtime coding assistants и streaming UX.
US gov получает до 30 days pre-release access для security review. Non-binding on paper, effective in practice. 26 June: OpenAI cap at ~20 trusted partners (OSTP + ONCD coordination).
| Lab | Model | Status |
|---|---|---|
| OpenAI | GPT-5.6 Sol/Terra/Luna | Preview ~20 partners |
| Anthropic | Claude Fable 5 / Mythos 5 | 12 June: export kill switch, global offline |
| Gemini 3.5 Pro | Slipped to July (was June) |
| Dimension | GPT-5.6 Sol | Claude Mythos 5 |
|---|---|---|
| TerminalBench 2.1 | 91,9% (Ultra) / 88,8% | 88,0% |
| ExploitBench | ≈ Mythos Preview, ~⅓ tokens | No public data |
| Input price | $5 / M | $10 / M (offline) |
| Availability | Limited preview, GA in weeks | Export control — offline |
| Context window | ~1.5M tok | 200K tok |
Sol wins coding/cyber benchmarks при 2× cheaper input. Fable 5 still competitive на SWE-bench Pro — full compare после System Card drop. Background: Claude Fable 5 export control breakdown.
Watch OpenAI status page: alert на public API unlock
Hold production baseline: GPT-5.5 или Claude Opus 4.8 до GA
Pre-map model routing: Sol → agents, Terra → mass API, Luna → lightweight
Priority eval post-GA: TerminalBench-style pipelines, CTF research, long-context doc RAG
Model token cost: Ultra только для genuinely hard tasks — burn rate spikes
Cerebras ROI calc: post-July enterprise channel для 750 tok/s eval
| Requirement | Pick |
|---|---|
| Complex codegen, debug, multi-agent tasks | Sol |
| Enterprise doc analysis, support, mass API | Terra |
| High-QPS summary, drafting, automation | Luna |
| GPT-5.5-tier on tight budget | Terra (−50% cost) |
| Latency-critical realtime (post-July) | Sol on Cerebras |
Pure cloud API = fast model swap, но policy whiplash, long-context bill shock и unpredictable Ultra token burn. Self-host = A100/H100 capex + ops headcount. Для 7×24 AI agents, multi-agent coding pipelines или iOS CI/CD automation в production — NodeMini Mac Mini M4 cloud rental даёт unified memory + Apple Silicon efficiency как stable execution layer. Pricing: тарифы аренды.
Для public users — нет. ~20 trusted partners через API/Codex. ChatGPT rollout ожидается July 2026; Polymarket: 87% до 31 July.
TerminalBench: 91,9% (Ultra) vs 88,0%. ExploitBench parity при ~⅓ tokens. Mythos 5 силён на SWE-bench Pro. См. export control analysis.
Multi-agent orchestration: decompose → parallel sub-agents → merge. Driver 91,9% TerminalBench; token burn выше — reserve для hard workloads only.
Executive order 2 June 2026 → OSTP/ONCD security review gate. OpenAI capped at ~20 partners; CEO против long-term normalization этого pattern.
С July 2026: до 750 tok/s на GPT-5.6 Sol — 5–15× vs типичные 50–150 tok/s. Initial rollout: selected enterprise only.
Sol: complex coding + multi-step agents. Terra: enterprise docs + mass API. Luna: summary + automation. Execution layer: help center или coding assistants shootout.