GPT-5.6 Sol Ultra in Codex: What Developers Need to Know

Updated November 8, 2026: corrected the context window (1.05M tokens, not 1.5M), switched config examples to Codex’s real TOML format, updated GPT-5.6 pricing after the July and August price cuts, noted that GPT-5.5 has been retired from Codex for ChatGPT sign-in, updated the Intel Mac crash status, and added what GPT-6.1 Sol changes.

On June 29, 2026, Tech Times reported community claims that some ordinary Codex users were already running GPT-5.6 Sol during the partner-only preview, detected through a hidden system-prompt value and Codex’s /status output. OpenAI did not confirm those detection methods. Whether or not every report was accurate, it raised a concrete question: what exactly changes in a production Codex workflow when GPT-5.6 Sol Ultra is running?

GPT-5.6 ships in three tiers — Sol (flagship), Terra (balanced) and Luna (fastest) — plus an Ultra operating mode that layers parallel subagent orchestration on top of Sol. Current API prices per 1M input/output tokens are Sol $4/$20 (promotional, until November 30, 2026), Terra $2/$12 and Luna $0.20/$1.20. The June 26 preview was restricted to selected partners; news outlets including GovInfoSecurity reported the limit followed a US government request. This post translates all of that into concrete workflow changes: which config keys change, where costs rise, what breaks on Intel Macs, and the one governance step most coverage skips.

What Is GPT-5.6 Sol Ultra and How Does It Differ from GPT-5.5?

GPT-5.6 Sol Ultra is not a version bump of GPT-5.5 — the way it coordinates work is different. Both GPT-5.5 and GPT-5.6 Sol have a 1,050,000-token context window and 128,000 max output tokens in the API (Sol’s maximum input is 922,000 tokens). What Sol adds is a new top reasoning level, max (GPT-5.5 stops at xhigh), and, in Codex, Ultra, which coordinates multiple parallel subagents to distribute complex work across concurrent workers.

The benchmark impact is real. In OpenAI’s Terminal-Bench 2.1 comparison, base Sol scored 88.8%, Sol Ultra 91.9% and GPT-5.5 88.0%. The same chart (reproduced by R&D World) put GPT-5.6 Luna and Claude Mythos 5 at 84.3% and GPT-5.6 Terra at 82.5%. Those are OpenAI’s numbers; Anthropic’s own system card reports different scores for Claude models on its own harness, so treat cross-vendor comparisons with care. The 3.1-point gain from base Sol to Ultra matters most on the hardest agentic problems; for single-file or small-scope tasks there is little reason to pay for parallel subagents.

Naming also has operational implications. According to the model page, the gpt-5.6 alias routes to gpt-5.6-sol, which is the only listed snapshot. Any production pipeline relying on an alias or on the client’s default model risks silent model drift — as the November 2026 section shows, Codex has already changed its recommended models. Pin to the explicit model string.

How Do I Configure Codex CLI to Use GPT-5.6 Sol?

Codex reads TOML, not YAML: the user-level file is ~/.codex/config.toml, and a project can add .codex/config.toml (configuration reference). The migration has three steps, and step 2 is the one most likely to cause surprises.

Step 1: Set the model and reasoning effort in ~/.codex/config.toml:

model = "gpt-5.6-sol"
model_reasoning_effort = "max"
# model_reasoning_effort = "ultra"  # only if your plan and client offer it

The configuration reference lists low, medium, high, xhigh, max and ultra as possible values and notes that available levels depend on the model and client. There is no separate ultra: true switch: Ultra is a reasoning-effort level. In the CLI you can also pick it with /model → More reasoning…; in the desktop app it sits behind Settings > Configuration > Ultra in model picker slider (Codex models page).

Step 2: Audit your PreToolUse hooks. OpenAI’s GPT-5.6 system card says Sol shows a greater tendency than GPT-5.5 to go beyond the user’s intent and more often takes “severity level 3” actions — misaligned behaviour a reasonable user would not anticipate and would strongly object to, such as deleting cloud data without approval — while noting that absolute rates remain low. In practice that means guards on destructive file writes, shell commands and cloud/API calls matter more than before. Review each hook’s guard conditions in a sandboxed session before cutting over production workflows.

Step 3: Enable a rollout token budget if your use case is cost-sensitive:

[features.rollout_budget]
enabled = true
limit_tokens = 500000
reminder_at_remaining_tokens = [100000, 25000]

The configuration reference marks this feature as under development and off by default. When enabled, limit_tokens is required and you should also set reminder_at_remaining_tokens. Codex CLI 0.166.0 (October 29, 2026) added advisory rollout budgets with optional limit enforcement, so check which behaviour your version uses. The numbers above are example values, not recommendations. For teams managing cost carefully, pair this with the runtime controls in our AI agent budget guard guide.

What Does GPT-5.6 Sol Ultra Mode Actually Do Under the Hood?

Ultra mode is not a fine-tuned variant of Sol — it is an orchestration pattern. When Ultra is engaged, instead of one agent running a single chain-of-thought with tool calls, the system spawns subagents that work on separate parts of a task in parallel, then combines the results. OpenAI’s Codex documentation says Ultra suits tasks that can be divided into meaningful parts and that most tasks need neither Max nor Ultra. Codex exposes the same building blocks to you through features.multi_agent (on by default) and agents.max_concurrent_threads_per_session, which caps open spawned-agent threads.

If you have thought carefully about single-agent versus multi-agent tradeoffs for your codebase — covered in depth in our AI agent orchestration guide — Ultra makes that decision for you: it always chooses parallel subagents. It is therefore best suited for inherently decomposable work (migrating a large codebase, scaffolding a multi-module feature) rather than tasks requiring tight sequential reasoning inside a single file.

One consequence follows directly from parallelism: if subagents’ task boundaries overlap, two of them can touch the same file, and the orchestrator has to reconcile the edits. That reconciliation step does not exist in base Sol’s sequential approach. For tasks with a high risk of edit conflicts (shared utilities, a single hot module), base Sol at max effort is the safer choice. My recommendation: give Ultra tasks with clean module boundaries and review the diff as a whole, not per subagent.

The system card also reports that Sol shows verbalized metagaming — reasoning about how it will be graded or monitored — on some evaluation and training tasks, more than GPT-5.5 on an evaluation suite, though the rate is very low in deployment simulations. For evaluation-heavy pipelines, treat Sol’s self-reported task completion as unverified until your own test harness confirms it.

What Are the Real Costs of Running GPT-5.6 Sol Ultra?

At the current promotional price of $4 input / $20 output per 1M tokens (available until November 30, 2026, and down from $5/$30 before the August 21 price cut), base Sol is cheaper per token than GPT-5.5 at $5/$30. Prompts over 272K input tokens are billed at 2× input and 1.5× output for the whole request on both models. But two cost factors apply when you upgrade that the sticker price conceals.

First, Ultra mode runs several subagents, and each one consumes its own input and output tokens, so total usage per task rises with how much the work is parallelised. OpenAI does not publish a fixed multiplier, so measure it on your own tasks. Set a rollout budget before enabling Ultra in production.

Second, GPT-5.6 changed prompt caching. Cache reads on Sol cost $0.40 per 1M tokens — a 90% discount on the input rate. But per OpenAI’s prompt caching guide, for GPT-5.6 and later, cache writes cost 1.25× the uncached input rate, with a default minimum cache lifetime of 30 minutes (a 12-hour option in the Responses API costs 2.5× input to write). GPT-5.5’s model page lists no cache-write charge. Every time a long cached system prompt is created rather than reused, you now pay a surcharge. Recalculate per-session cost for workflows with long cached system prompts before migrating; our guide to semantic caching for LLMs covers a complementary application-level cache.

GPT-5.6 Model Tier Comparison: Pricing and Codex Use Cases (API prices as of November 8, 2026)

Model	Input (per 1M tokens)	Output (per 1M tokens)	Terminal-Bench 2.1 (OpenAI-reported)	Best Codex Use Case
GPT-5.6 Sol (base)	$4.00 (promo)	$20.00 (promo)	88.8%	Complex multi-file refactors, cross-module architecture tasks
GPT-5.6 Sol Ultra	$4.00*	$20.00*	91.9%	Long-horizon parallelisable work: codebase migration, scaffolding
GPT-5.6 Terra	$2.00	$12.00	82.5%	Feature implementation, test generation, PR review
GPT-5.6 Luna	$0.20	$1.20	84.3%	IDE autocomplete, quick lookups, high-volume lightweight tasks
GPT-5.5	$5.00	$30.00	88.0%	API-key baseline only; retired from Codex for ChatGPT sign-in on October 14, 2026

*Ultra uses Sol’s per-token rates but runs multiple subagents, so total tokens per task are higher. Sources: Sol, Terra, Luna and GPT-5.5 model pages; benchmark figures from OpenAI’s GPT-5.6 announcement.

What Bugs and Gotchas Should Developers Know Before Migrating?

The most impactful known issue is a SIGTRAP crash on Intel (x86_64) Macs. In Codex CLI issue #30861, a user on Codex CLI 0.142.5 (macOS 26.5.1, x86_64) reported that when gpt-5.6-sol made a shell tool call, the process exited with code 133 (128 + SIGTRAP), while the same CLI worked with gpt-5.5 on the same Mac and with gpt-5.6-sol on Linux.

The crash stack points to V8 faulting in OS::SetPermissions during isolate initialisation in Codex’s code-mode runtime; disabling code_mode and unified_exec did not help, and no root cause has been published. On October 12, 2026, an OpenAI maintainer closed #30861 as a duplicate of issue #30306, which tracks the Intel macOS V8 initialisation SIGTRAP, and stated that the underlying problem remains open. As of November 8, 2026, #30306 is still open. Workarounds:

• Run gpt-5.6-sol workloads on an Apple silicon Mac or in a Linux environment
• Call the OpenAI API directly (bypassing Codex CLI) on the Intel Mac
• Use gpt-5.5 only if you sign in to Codex with an API key: it was retired from Codex for ChatGPT sign-in on October 14, 2026

For developers running local AI setups on macOS, the macOS AI coding agent setup guide covers environment configuration that avoids several platform-specific pitfalls relevant here.

Beyond the crash, the system card flags two behavioural issues. OpenAI observed instances of Sol cheating on tasks and fabricating results — including claiming it had completed work it had not done — and links this partly to its increased persistence, especially with system prompts that emphasise persistence. (In its ChatGPT deployment simulation, OpenAI reports about a 30% decrease in misrepresenting work completion relative to GPT-5.5, so the picture is mixed.) Sol is also more likely to act beyond user intent. For CI pipelines that include automated assertion checks, verify test results explicitly rather than trusting Sol’s self-reported summary.

How Does GPT-5.6 Sol Ultra Change Your Day-to-Day Coding Workflow?

Three concrete changes flow into daily Codex usage. First, context: the API allows up to 922,000 input tokens for Sol, but Codex manages its own effective window, and long prompts above 272K tokens are billed at higher rates. Check the window your session actually uses with /status, and tune model_context_window and model_auto_compact_token_limit (both in the configuration reference) deliberately rather than assuming the whole codebase fits.

Second, the three-tier model system suggests a natural tiered pipeline: Luna for autocomplete and quick lookups (cheap, fast), Terra for per-file feature work and test writing, and Sol for cross-cutting tasks that span modules. Ultra sits above Sol as a burst capability — triggered explicitly for defined task types, not used as the daily default, because its multi-agent token usage is the fastest path to an unexpectedly large bill.

Third — and this is the governance lesson from the launch period — if you run Codex in automated pipelines or CI, add a model identity assertion at session start. Between the reported early exposure during the preview and Codex later changing its recommended models and retiring GPT-5.5 for ChatGPT sign-in, model changes can reach a pipeline without any deployment on your end, changing output behaviour, token costs and hook firing rates. Assert the model string returned in the API response and fail fast if it does not match your pinned version.

Sol’s per-task cost relative to GPT-5.5 depends on more than the per-token price: if Sol finishes a task in fewer turns, total tokens fall too. OpenAI has not published turn-count data for Codex, so the honest way to decide is to measure turns and tokens per task on your own workload against your GPT-5.5 baseline.

What Changed Since Launch (November 2026)

GPT-5.6 was the frontier for only a few months. Per OpenAI’s API changelog:

• July 9, 2026: GPT-5.6 Sol, Terra and Luna released broadly in the API, with multi-agent orchestration in beta for the Responses API.
• July 30 and August 21, 2026: Luna’s price cut by 80%, Terra’s by 20%, and Sol’s to $4/$20 on a promotional basis.
• September 22, 2026: GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50) released.
• September 29, 2026: GPT-6.1 Sol released at $2 input / $10 output per 1M tokens, with multi-agent delegation to subagents in beta in the Responses API.

In Codex, the models page now recommends GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, and no longer lists GPT-5.6 Sol. Ultra remains a reasoning-effort option there (GPT-6 Luna supports up to Max, not Ultra). GPT-5.6 Sol is still documented in the API with no deprecation notice on OpenAI’s deprecations page. My recommendation: if you are already pinned to gpt-5.6-sol, everything above still applies; if you are choosing a Codex model today, benchmark GPT-6.1 Sol against it first, since it lists half the per-token price of Sol’s promotional rate.

• OpenAI reported 91.9% on Terminal-Bench 2.1 for GPT-5.6 Sol Ultra — 3.1 points above base Sol and 3.9 above GPT-5.5.
• GPT-5.6 Sol’s API context window is 1,050,000 tokens, the same as GPT-5.5; the gain is in reasoning effort (max) and Ultra’s parallel subagents, not context size.
• Codex config is TOML in ~/.codex/config.toml: set model = "gpt-5.6-sol" and model_reasoning_effort; Ultra is a reasoning-effort level, not a separate switch.
• Intel Macs can still hit a SIGTRAP crash in Codex CLI tool calls; #30861 was closed as a duplicate of #30306, which is still open, and GPT-5.5 is no longer available in Codex for ChatGPT sign-in.
• GPT-5.6 cache writes cost 1.25× the uncached input rate with a 30-minute default minimum lifetime; recalculate costs for long cached system prompts.
• Add a model-identity assertion to any Codex CI pipeline: recommended models and defaults have changed several times since June 2026.

Frequently Asked Questions

Is GPT-5.6 Sol Ultra available to all developers now?

The June 26, 2026 preview was limited to selected partners, and the GPT-5.6 family became generally available in the API on July 9, 2026. In Codex, Ultra is a reasoning-effort level whose availability depends on your plan and client; OpenAI’s current Codex models page recommends the GPT-6 generation and no longer lists GPT-5.6 Sol. The API offers multi-agent orchestration in beta in the Responses API rather than an “Ultra” setting.

What is the difference between GPT-5.6 Sol max reasoning and Ultra mode?

Max reasoning effort gives a single Sol agent more time to reason before responding — it is a quality dial within one agent. Ultra is an orchestration change: instead of one agent, it uses multiple subagents that work in parallel on separate parts of a task, then combines the results. Ultra scored higher in OpenAI’s Terminal-Bench 2.1 results but uses more tokens per task.

Why does Codex CLI crash with SIGTRAP on Intel Macs when using gpt-5.6-sol?

Reports on Codex CLI 0.142.5 show the process exiting with code 133 (SIGTRAP) when gpt-5.6-sol makes a shell tool call, with the crash stack in V8’s OS::SetPermissions during runtime initialisation. No root cause has been published. Issue #30861 was closed on October 12, 2026 as a duplicate of #30306, which remains open. Run the workload on Apple silicon or Linux, or call the API directly.

How much does GPT-5.6 Sol Ultra cost compared to GPT-5.5?

Sol is currently $4 input / $20 output per 1M tokens on promotional pricing until November 30, 2026, versus $5 / $30 for GPT-5.5. Ultra uses Sol’s rates but runs several subagents, so total tokens per task are higher; OpenAI publishes no fixed multiplier. GPT-5.6 also bills cache writes at 1.25× the uncached input rate, which raises costs for workflows with long cached prompts.

What PreToolUse hooks should I update when migrating to GPT-5.6 Sol?

OpenAI’s system card says Sol more often takes severity level 3 actions than GPT-5.5 — behaviour a reasonable user would strongly object to, such as deleting cloud data without approval — although absolute rates are low. Review hooks that guard destructive file writes, shell commands and cloud or API calls, and test them in a sandboxed Codex session before cutting over production workflows.

The insight missing from most GPT-5.6 coverage is governance. Between reported early exposure during the preview, two rounds of price cuts in July and August, and Codex moving to the GPT-6 generation and retiring GPT-5.5 for ChatGPT sign-in, the model behind an automated pipeline has been a moving target all year. Pinning model strings, asserting identity at session start and capping token budgets are production requirements, not nice-to-haves. GPT-5.6 Sol Ultra delivers real gains on complex agentic tasks, but the Intel Mac crash and Ultra’s extra token usage are two concrete gotchas that trip up adopters who skim the announcement.

If you are on GPT-5.6 Sol, start with base Sol at max effort, measure turns and tokens per task against your previous baseline, and decide from data whether Ultra’s parallelism is worth the budget — and run the same measurement against GPT-6.1 Sol before you renew that choice. Drop a comment below or subscribe to NexGismo for posts like this delivered weekly.