OpenAI Launches GPT-5.6 Series and ChatGPT Work: Coding Agent Index Hits 80 (SOTA), Ultra Tier Runs 4 Parallel Agents

Forum topic · 小凯 · 2026-07-11

Summary

On July 9, 2026, OpenAI released the GPT-5.6 model family (Sol, Terra, Luna) with two reasoning tiers (max and ultra), plus ChatGPT Work, an autonomous agent that works across apps and files for hours. GPT-5.6 Sol with max reasoning scored 80.0 on Artificial Analysis' Coding Agent Index v1.1, a new public SOTA, while using less than half the output tokens, time, and roughly one-third the cost of the runner-up. Pricing ranges from Luna ($1/$6 per 1M tokens) to Sol ($5/$30). The ultra tier coordinates 4 parallel agents by default, expandable to 16 via the Responses API. The release also introduces Programmatic Tool Calling, where the model writes and runs lightweight programs in memory to orchestrate tools and filter intermediate data, compatible with Zero Data Retention for regulated industries. ChatGPT Work builds on Codex, now used by 5 million weekly users, targeting cross-application enterprise workflows. The post also weighs benchmark-selection concerns, open questions about agent autonomy, and competitive moves from Cognition's SWE-1.7 and Meituan's LongCat-2.0.

On July 9, 2026, OpenAI announced two things simultaneously: the GPT-5.6 model series and ChatGPT Work. Source links:

Announcement: 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 Release: A Three-Tier Lineup Plus a Work Agent

Rather than a single model, OpenAI shipped three GPT-5.6 tiers (Sol / Terra / Luna) plus two reasoning tiers (max and ultra), covering the full price-performance curve.

| 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 at lower cost | | Luna | 1.00 | 6.00 | Extreme cost-efficiency for large-scale deployment |

Reasoning tiers:

max: longer reasoning time and exploration space
ultra: coordinates 4 parallel agents by default; via the Responses API multi-agent beta, up to 16 agents

The second launch, ChatGPT Work, is an agent that can autonomously gather information across apps and files, break complex projects into steps, and work for hours. It builds on Codex, which OpenAI says now has over 5 million weekly users.

Key Benchmarks: Coding Agent Index 80.0 Is a New 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 used less than half the output tokens, less than half the time, and roughly one-third the cost compared to Fable 5 — a rare full-dimensional Pareto frontier advance.

Other Sol results:

Agents' Last Exam: 53.6 (new high), +13.1 over Fable 5
Terminal-Bench 2.1: 88.8% / Sol Ultra 91.9% (SOTA)
BrowseComp: 90.4% / Sol Ultra 92.2% (SOTA)
OSWorld 2.0: 62.6%, beating Opus 4.8 with 85% fewer output tokens
SEC-Bench Pro: 71.2% / Sol Ultra 74.3%
ExploitBench 2: 73.5% (GPT-5.5: only 47.9%)
DeepSWE v1.1: 72.7%

Programmatic Tool Calling

The most architecturally significant piece of the release. Traditional agent flows either require developers to hand-orchestrate every step, or pass every tool response back to the model — the former is engineering-heavy, the latter explodes token consumption.

Programmatic Tool Calling lets GPT-5.6 write and run lightweight programs in memory to:

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

This shifts the agent loop from a serial "model thinks → tool call → model thinks" pattern to "model writes program → program calls multiple tools → program aggregates → model decides" — a parallel pattern. It is also Zero Data Retention (ZDR) compatible, removing a barrier for financial, healthcare, and government customers with strict data-residency requirements.

ChatGPT Work: Codex Goes Cross-App

ChatGPT Work is positioned as an agent that can gather information across apps and files, decompose complex projects into independently completed steps, and keep working for hours, powered by GPT-5.6. Codex's evolution from an IDE coding assistant into a cross-application digital employee — reading Slack, Notion, Excel, GitHub, Confluence — shifts the purchase decision from tooling budgets (developer spend) to labor budgets (operations spend), a key path for OpenAI raising enterprise ARPU.

Why It Matters

1. Tiered models + tiered reasoning is a maturing product strategy

Three tiers × two reasoning modes = 6 capability-cost combinations. Users pay for their actual workload (Luna for document summarization, Sol + ultra for complex multi-step research). For competitors, this creates direct product pressure: Anthropic and Google must either match the tiered approach or compete on single-model price.

2. What 80 on the Coding Agent Index means

The index is a composite agentic-coding benchmark covering real engineering tasks. 80 means the model independently completes ~80% of end-to-end tasks (requirements → codebase reading → editing → testing → submission). The trajectory — ~30% in 2024, ~60% in 2025, 80% in 2026 — makes the "AI replaces junior programmers" forecast a quantifiable reality.

3. The Chinese angle

The same day, Cognition released SWE-1.7 (built on Kimi K2.7) scoring 81.5% on Terminal-Bench, and Meituan's LongCat-2.0 surpassed GPT-5.5 on SWE-bench Pro (reported 07-06). Frontier AI coding capability is no longer a single-company or single-country game. Positive signal for China: Kimi K2.7 as an internationally adopted base model. Negative: core increments (RL pipelines, agent harnesses, product experience) are still mostly defined by US companies.

Risks and Open Questions

1. Benchmark-selection concerns: Artificial Analysis' eval sets and weights are public, so vendors can optimize for them. SWE-1.7 leads slightly on Cognition's in-house FrontierCode (42.3%) but trails GPT-5.6 Sol by multiple points on the Coding Agent Index — expect more "same capability, different benchmark, different ranking" situations. 2. Is ChatGPT Work truly autonomous or template-driven? The announcement doesn't disclose its mechanics. Preset workflow templates would make it an advanced automation tool; an open-ended agent loop would be another order of product. 3. Does 4-agent parallelism in ultra actually pay off? Coordination overhead may eat parallel gains, and OpenAI hasn't published end-to-end latency data. 4. Sol's $5/$30 pricing isn't cheap. Luna's $1 input is the first sub-dollar frontier-tier option, but whether high-throughput inference can sustain the Sol 80-point experience remains open.

Overall, GPT-5.6 + ChatGPT Work is OpenAI's most important product move of mid-2026: tiered models covering all user scenarios, Codex moving from tool to employee, and a bid for pricing power in the enterprise market. Anthropic and Google must respond within the next 12 months.