GPT-5.6 Sol: The Codex Merger Has Arrived

July 9, 2026, was a strange day for AI observers. Most people expected OpenAI to ship a new model. The GPT-5.6 family had been in limited preview since June 26 after all. What nobody predicted was a product restructure this sweeping. OpenAI didn't just release a model. It collapsed its standalone Codex coding agent into the ChatGPT desktop app, launched a new agentic mode called ChatGPT Work, and shipped three model tiers (Sol, Terra, and Luna) at different price points. The old ChatGPT desktop app is now called "ChatGPT Classic." Codex is no longer a separate download.

This is not a model update. It is a platform consolidation. And it changes how developers and knowledge workers should think about OpenAI's product stack.

The Codex Merger: What Actually Changed

Codex is not shutting down. It is being absorbed. The Codex desktop app becomes the new ChatGPT desktop app, and Codex survives as a dedicated coding mode inside that unified surface. OpenAI confirmed that "Codex keeps its dedicated coding experience alongside Chat and Work, with inline editing in diffs, pull request review in the side panel, faster Computer Use powered by GPT-5.6, and multi-repository projects."

The new desktop app has three modes:

• Chat: The familiar ChatGPT conversation interface, now in a sidecar window.
• Work: A new agentic mode powered by GPT-5.6, designed for longer cross-app knowledge work. Drafting documents, building spreadsheets, operating browsers autonomously.
• Codex: The same coding agent developers have been using, with the same project system and plugin directory.

The merger has been brewing for months. Starting with GPT-5.4 in April 2026, OpenAI merged the Codex model line into the main GPT model, discontinuing Codex as a separately iterated model family. The desktop app merger was the inevitable product-side follow-up. Greg Brockman took charge of the unified product organization in May, writing in an internal memo that OpenAI would "invest in a single agentic platform."

For the 5 million weekly Codex users OpenAI cited at launch, virtually nothing changes. Update your app and it becomes the new ChatGPT desktop app. You can even keep the Codex icon.

GPT-5.6 Model Family: Sol, Terra, Luna

The model family comes in three tiers. The naming (Sol, Terra, Luna) signals a deliberate hierarchy. OpenAI calls them "durable capability tiers that can advance on their own cadence."

Tier	Input Price (per 1M tokens)	Output Price (per 1M tokens)	Context Window	Max Output	Best For
GPT-5.6 Sol	$5.00	$30.00	1,050,000	128,000	Frontier reasoning, agentic coding, research
GPT-5.6 Terra	$2.50	$15.00	1,050,000	128,000	Balanced work, competitive with GPT-5.5
GPT-5.6 Luna	$1.00	$6.00	1,050,000	128,000	High-volume, cost-sensitive tasks

The gpt-5.6 alias routes to Sol. All three tiers support the 1.05M token context window and 128K max output tokens. Those are the same top-end specs OpenAI first introduced with GPT-5.5.

OpenAI also introduced two new reasoning modes for Sol: max (deeper reasoning, similar to Anthropic's extended thinking) and ultra (leverages sub-agents that split and parallelize complex work). During the preview, OpenAI also announced that Sol would run on Cerebras hardware at up to 750 tokens per second, starting in July 2026.

Prompt caching got an upgrade: explicit cache breakpoints, a guaranteed 30-minute minimum cache life, and cache writes billed at 1.25x the uncached input rate. Cache reads still get the 90% discount.

Coding Performance: The New State of the Art (With Caveats)

Coding is where GPT-5.6 Sol makes its strongest case, and also where the picture gets complicated. OpenAI published an extensive benchmark suite at GA. Third-party evaluator Artificial Analysis published independent results on July 9.

On the Artificial Analysis Coding Agent Index (an independent index pairing models with agentic harnesses across DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA), Sol (max) leads at 80. That is ahead of Claude Fable 5 at 77.2 and GPT-5.5 at 76.4. Terra (77.4) also edges Fable 5. Luna (74.6) outperforms Opus 4.8 (72.5). The independent numbers broadly confirm OpenAI's claims.

But the SWE-Bench Pro story is different. OpenAI's own comparison table shows Sol at 64.6% while Claude Mythos 5 reaches 80.3% and Fable 5 hits 80%. That is a 15-point gap, and it is the biggest weakness in Sol's coding profile.

Benchmark	GPT-5.6 Sol	GPT-5.6 Sol Ultra	GPT-5.6 Terra	GPT-5.6 Luna	Claude Fable 5	Claude Mythos 5	Claude Opus 4.8
AA Coding Agent Index	80	--	77.4	74.6	77.2	--	72.5
Terminal-Bench 2.1	88.8%	91.9%	87.4%	84.7%	83.1%	88.0%	78.9%
SWE-Bench Pro	64.6%	--	63.4%	62.7%	80.0%	80.3%	69.2%
DeepSWE v1.1	72.7%	--	69.6%	67.2%	69.7%	--	59.0%

The pattern is clear. Sol dominates agentic terminal-based coding (Terminal-Bench, DeepSWE) but trails on repository-level issue resolution (SWE-Bench Pro). For builders, this means Sol is strong for greenfield development and multi-step agentic workflows. If your work involves patching large existing codebases from issue descriptions, Fable 5 or Mythos 5 may still be the better call.

Pricing: Half the Cost of Fable 5

OpenAI priced aggressively. Sol costs $5/$30 per 1M tokens, exactly half of Claude Fable 5 at $10/$50. Terra at $2.50/$15 matches GPT-5.5's performance at half the price. Luna at $1/$6 outperforms Opus 4.8 ($5/$25) at roughly one-quarter the cost.

Model	Input (per 1M)	Output (per 1M)	Total (1M in + 1M out)
GPT-5.6 Sol	$5.00	$30.00	$35.00
GPT-5.6 Terra	$2.50	$15.00	$17.50
GPT-5.6 Luna	$1.00	$6.00	$7.00
Claude Fable 5	$10.00	$50.00	$60.00
Claude Opus 4.8	$5.00	$25.00	$30.00
Gemini 3.1 Pro Preview	$2.50	$17.50	$20.00

Artificial Analysis calculated the cost per task on their Intelligence Index. Sol costs $1.04 per task at max reasoning, compared to roughly $3+ for Fable 5. On the Coding Agent Index, Sol's per-task cost is about 40% cheaper than Fable 5 in Claude Code, and about 10% cheaper than Opus 4.8.

For teams that route tasks by difficulty, the 5x price spread between Sol and Luna means you can use Luna for high-volume simple tasks and reserve Sol for the hardest problems, all under one API key.

Beyond Coding: Cybersecurity, Research, and Agents

OpenAI positions GPT-5.6 as its strongest cybersecurity model yet. On internal evaluations, Sol achieved frontier performance on ExploitBench with about 80% fewer output tokens than Claude Mythos 5. The system card classifies Sol, Terra, and Luna all as "High capability" in cybersecurity, backing this with roughly 700,000 A100e GPU hours of automated red-teaming.

On Agents' Last Exam (an evaluation of long-running professional workflows across 55 fields), Sol scores 53.6, beating Fable 5 by 13.1 points. Even at medium reasoning, Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.

BrowseComp, a benchmark for autonomous web research, sees Sol at 90.4% (92.2% with ultra). That is ahead of GPT-5.5 and Opus 4.8.

Early users have also praised Sol's writing and design capabilities. Claire Vo's "How I AI" vibe benchmark showed Sol winning across PRDs, prototypes, wireframes, and agentic voice. She also noted that taste-based benchmarks are inherently subjective.

Availability: Who Gets What

The rollout is tiered by subscription:

• ChatGPT Free and Go: Access GPT-5.6 Terra in ChatGPT Work and Codex
• ChatGPT Plus: Access Sol at medium and higher effort settings
• ChatGPT Pro and Enterprise: Access Sol Pro, plus can select effort per model
• API: All three tiers available via v1/responses, v1/chat/completions, and v1/batch

The government-mandated limited preview (June 26 to July 8) restricted Sol to about 20 organizations. OpenAI announced on July 8 that GA would proceed as planned on July 9, with CEO Sam Altman posting "Happy building."

Microsoft 365 Copilot immediately adopted GPT-5.6 as its preferred model, making it the default for Microsoft's enterprise AI suite.

Developer Experience: What Builders Should Know

OpenAI also introduced Programmatic Tool Calling. This lets the model write JavaScript that runs in an isolated V8 runtime with no network access. Early results show 38% to 63.5% token reductions for documented API contracts.

For developers already using Codex, the day-to-day workflow does not change. The same project system, plugin directory, and @-mentions work inside the unified app. What changes is the ceiling. GPT-5.6 Sol in Codex handles multi-repository projects, longer agentic sessions, and more complex tool chains than GPT-5.5 could.

The short version: if you build coding agents, Sol is now the cheapest path to a top-tier index score. If you patch enterprise repos, benchmark Fable 5 before switching. For everyone else, the 5x price spread across Sol/Terra/Luna makes this the most flexible model family OpenAI has shipped. And the Codex merger means you no longer have to choose between a coding tool and a chat tool.

---