> 2026-07-09 · OpenAI · AI Agent Commercialization > Original: https://openai.com/index/gpt-5-6/ > ChatGPT Work: https://openai.com/index/chatgpt-for-your-most-ambitious-work > Programmatic Tool Calling docs: https://developers.openai.com/api/docs/guides/tools-programmatic-tool-calling > System Card: https://deploymentsafety.openai.com/gpt-5-6
The Announcement: A Three-Tier Family Plus a Work-Grade Agent
On July 9, OpenAI announced two things at once:
1. The GPT-5.6 series in three tiers (Sol / Terra / Luna), each with two reasoning modes (max and ultra), covering the full price-performance curve from maximum intelligence to extreme cost efficiency.
2. ChatGPT Work, an AI agent that can work autonomously across apps and files for hours — "for your most ambitious work." With over 5 million weekly Codex users, ChatGPT Work is the evolution of Codex technology into cross-application workflows.
| Model | Input \(/1M | Output\)/1M | Positioning | |-------|-----------|-------------|-------------| | Sol | 5.00 | 30.00 | Flagship, dual SOTA in intelligence + efficiency | | Terra | 2.50 | 15.00 | Balanced, matches GPT-5.5 performance at lower cost | | Luna | 1.00 | 6.00 | Extreme value, mass deployment |
Each model has two reasoning modes:
- max: longer reasoning time and exploration budget
- ultra: coordinates 4 agents in parallel by default; via the Responses API multi-agent beta, up to 16 agents can be configured
- Agents' Last Exam: Sol 53.6 (new record) — 13.1 points above Fable 5
- Terminal-Bench 2.1: Sol 88.8% / Sol Ultra 91.9% (SOTA)
- BrowseComp: Sol 90.4% / Sol Ultra 92.2% (SOTA)
- OSWorld 2.0: Sol 62.6%, beating Opus 4.8 with 85% fewer output tokens
- SEC-Bench Pro: Sol 71.2% / Sol Ultra 74.3%
- ExploitBench 2: Sol 73.5% (GPT-5.5 only 47.9%)
- DeepSWE v1.1: Sol 72.7%
- Orchestrate tool calls
- Process intermediate results
- Monitor progress
- Dynamically choose next actions
- Filter large volumes of intermediate data, keeping only key information, drastically reducing tokens returned to the model
- Collects information across apps and files
- Decomposes complex projects into small steps
- Completes them independently, without continuous human intervention
- Works for hours at a stretch
- Powered by GPT-5.6, with multi-step reasoning and template-based material generation
Key Benchmark: Coding Agent Index 80 Sets SOTA
GPT-5.6 Sol (max reasoning) scored 80.0 on Artificial Analysis' Coding Agent Index v1.1 — the current public SOTA.
| Model | Coding Agent Index v1.1 | |-------|--------------------------| | GPT-5.6 Sol (max) | 80.0 (SOTA) | | GPT-5.6 Terra | 77.4 | | Claude Fable 5 | 77.2 | | GPT-5.5 | 76.4 | | GPT-5.6 Luna | 74.6 | | Claude Opus 4.8 | 72.5 | | Gemini 3.1 Pro Preview | 42.7 |
Notably, Sol uses less than half the output tokens, less than half the time, and roughly one-third the cost of Fable 5 — intelligence improved while token, time, and monetary costs all fell. This is a rare full-dimensional Pareto frontier advance.
Other Sol results:
Programmatic Tool Calling: The Model Writes Code to Orchestrate Tools
This is the most architecturally significant part of the release.
Traditional agent flows require developers to hand-orchestrate every step, or feed every tool response back to the model for next-step decisions — the first is engineering-heavy, the second explodes token consumption.
Programmatic Tool Calling lets GPT-5.6 write and run lightweight programs in memory to:
The engineering value: it advances the agent loop from the serial "model thinks → tool call → model thinks again" pattern to a parallel "model writes program → program calls multiple tools → program aggregates results → model decides" pattern.
It ships with Zero Data Retention (ZDR) compatibility, making it suitable for customers with strict data-residency requirements (financial institutions, healthcare, government) — removing one barrier that previously blocked agent adoption in sensitive industries.
ChatGPT Work: Taking Codex Cross-Application
ChatGPT Work's positioning is clear: an AI agent that gathers information across apps and files, breaks complex projects into small steps, completes them independently, and keeps working for hours.
OpenAI reports over 5 million weekly Codex users — a signal that Codex is no longer a geek toy but a mainstream developer tool.
ChatGPT Work's capability stack:
Why It Matters
1. Tiered models + tiered reasoning modes is maturing product strategy
GPT-5.6 is not one model but three tiers × two reasoning modes = 6 capability-cost combinations. Users pay for actual needs rather than full frontier capability — document summarization on Luna, complex multi-step research on Sol + ultra. For OpenAI, this is finer-grained token economics. For competitors, it creates direct pressure: Anthropic and Google must either match the tiering or compete on a single model at lower prices.
2. What an 80 on the Coding Agent Index means
Artificial Analysis' index is a composite agentic coding benchmark spanning real engineering tasks. 80 means the model independently completes 80% of end-to-end tasks (read requirements → explore codebase → modify code → test → commit). We're at the threshold from the "Copilot era" to the "Agent era": 5 years ago SWE-Bench was 0%, 2024 was 30%, 2025 was 60%, 2026 is 80%. For the software industry, "AI replacing junior programmers" has shifted from hypothesis to quantifiable reality — contextualized by Stanford's data (July 7) showing a 19% employment decline among 22–25-year-old software developers.
3. ChatGPT Work pushes Codex from tool to cross-app employee
Codex was an IDE coding assistant; ChatGPT Work positions it as a cross-application digital employee reading Slack, checking Notion, querying Excel, editing GitHub, and writing Confluence — not just writing code. In enterprise purchasing, "AI that writes code" (tool budget) and "AI that completes projects" (labor budget) are entirely different decisions. ChatGPT Work targets the latter — a key path to raising enterprise ARPU.
4. The China angle
The same day GPT-5.6 set the Coding Agent Index SOTA, Cognition released SWE-1.7 (built on Kimi K2.7) at 81.5% on Terminal-Bench, and Meituan's LongCat-2.0 reportedly beat GPT-5.5 on SWE-bench Pro (July 6). Frontier AI coding is no longer a single-company, single-country game. For China's AI industry, this is a two-sided signal: positively, Kimi K2.7 being chosen as Cognition's base model shows international recognition of Chinese open-source engineering; negatively, core frontier increments (RL pipelines, agent harnesses, product experience) are still defined mainly by US companies.
Risks and Open Questions
1. Benchmark selection controversy: Artificial Analysis' benchmark sets and weights are public, so vendors can optimize for them. SWE-1.7 scores 42.3% on Cognition's own FrontierCode (slightly ahead of GPT-5.5) but trails Sol on Artificial Analysis' index. Expect "same capability, different benchmarks, different rankings" to become more common.
2. Is ChatGPT Work truly autonomous or template-driven? The announcement says it "breaks projects into steps and completes them independently" but doesn't disclose the mechanism. Preset workflow templates would make it an advanced automation tool; a genuine open agent loop would be a different order of product.
3. Does 4-agent parallelism actually pay off? Coordination overhead can eat parallel gains. OpenAI hasn't published end-to-end latency data for the ultra tier.
4. Sol's $5/$30 pricing isn't cheap. Luna's $1/$6 is the first time a frontier model's lightweight tier dropped to $1 input. Whether Cerebras-class inference speed at 0.25x cost can support the Sol-tier experience remains open.
Overall, GPT-5.6 + ChatGPT Work is OpenAI's most important release of mid-2026, with a clear strategic intent: cover every user scenario with tiered models, evolve Codex from tool to employee, and establish pricing power in the enterprise market. Anthropic and Google must respond within the next 12 months.