Claude Opus 5.5 vs GPT-6 Sol: Coding Agent Index Score Tested (September 2026)
Anthropic and OpenAI both shipped frontier updates on the same day, September 22, 2026: Claude Opus 5.5 became Claude Code's default model, and GPT-6 Sol landed as the general access tier of OpenAI's new split release alongside the more gated GPT-6 Luna. In the September 24 refresh of the Artificial Analysis Coding Agent Index, Claude Code on Opus 5.5 scored 66 against Codex on Sol at 57, with Sol pricing Codex tasks at 2.99 dollars. That is a nine point gap on a shared, independently run index, and it is worth understanding what it measures before anyone treats it as a full verdict on which agent to standardize on.
Claude Opus 5.5 vs GPT-6 Sol: Coding Agent Index Score Tested (September 2026)
Anthropic and OpenAI picked the same day to move. On September 22, 2026, Anthropic shipped Claude Opus 5.5 and made it the default model inside Claude Code the same afternoon. OpenAI shipped GPT-6 Sol and GPT-6 Luna, a split release where Sol is the generally available tier and Luna sits behind a more gated, security focused access tier. Two days later, the Artificial Analysis Coding Agent Index refreshed with both models scored against the same harness: Claude Code running Opus 5.5 posted a 66, Codex running Sol posted a 57, and Codex on Sol now prices out at 2.99 dollars per completed task.
A nine point gap on a shared, third party index is a real signal, not noise. It is also a narrower claim than "Opus 5.5 is the better model," and the difference matters if you are deciding which agent your team standardizes on this quarter.
What actually shipped, and what the index measures
The Coding Agent Index does not score a raw model. It scores an agent, meaning a specific model paired with a specific harness, tool access pattern, and orchestration layer, run against a fixed battery of real coding tasks and averaged into one number. Claude Code with Opus 5.5 and Codex with Sol are two different products being compared, not two checkpoints of the same scaffolding swapped in and out. That distinction is easy to lose in a single headline number, and it is the first thing worth holding onto.
| Claude Code (Opus 5.5) | Codex (GPT-6 Sol) | |
|---|---|---|
| Release date | September 22, 2026 | September 22, 2026 |
| Coding Agent Index score (Sept 24 refresh) | 66 | 57 |
| Access tier | General availability, default model | General availability tier of a split release (Luna is gated) |
| Reported task cost | Not separately broken out in the refresh | 2.99 dollars per task |
| Prior generation comparison point | Opus 5 held the top spot before this update | GPT-6 Astra was the prior top OpenAI entry |
The scale of the move is also worth sitting with. Claude Code held the number one ranking before this refresh too, so Opus 5.5 extended a lead rather than overturning a standing. Codex moving to Sol kept it in second place rather than displacing Claude Code, which is a different outcome than a straight model upgrade "winning" the leaderboard outright.
Why the split release matters more than the score
GPT-6 Sol and GPT-6 Luna are not simply a fast tier and a slow tier of the same capability. The pattern across September's frontier releases, and OpenAI's Sol and Luna split is one clear example of it, is a separation between how capable a model is and what it is permitted to do unsupervised. Three of the month's four major frontier moves shipped a generally available model alongside a more gated, higher trust tier rather than a single unified release. Sol is the one Codex users get by default; Luna sits behind additional access controls that OpenAI has not fully detailed for general developer use as of this week.
That has a direct practical consequence for teams evaluating Codex against Claude Code on the index: the 57 you see is Sol's number, not necessarily the ceiling of what GPT-6 can do in an agentic coding harness. If Luna eventually becomes available to a broader developer tier, the gap could close, widen, or become irrelevant to whichever access level your organization actually qualifies for. Benchmarking the tier you can buy today is the right instinct, but it is worth building in the assumption that this number has a shorter shelf life than usual, because the vendor itself has signaled a second, more capable tier is coming.
What the nine point gap does and does not tell you
The Coding Agent Index blends performance across a range of task types, so a nine point gap is an average, not a category sweep. It is consistent with what has shown up on other shared benchmarks this year, where Claude Code has generally led on long horizon, multi step engineering tasks that require sustained tool use across many turns, while Codex has been competitive on narrower, single file completion work at a lower cost per task. Nothing in the September 24 refresh breaks that pattern; it mostly confirms it at new model versions.
The 2.99 dollar task cost for Codex on Sol is the number worth putting next to the score rather than reading in isolation. A team running high volume, lower complexity tasks, ticket triage style fixes, small refactors, test generation, may find that a lower per task cost on Sol offsets a lower aggregate index score, because the tasks that make up most of that volume are exactly the kind Codex has historically handled well relative to its price. A team leaning on an agent for long, ambiguous, multi file changes is optimizing for a different variable, and the index score is closer to directly predictive there.
What to actually do with this number this week
Do not switch your default coding agent off a single index refresh two days after both models shipped. Benchmarks this fresh have not absorbed a full week of real workload variance, and neither vendor has published a detailed task level breakdown of the September 24 run.
Do treat the release timing as a signal to re run your own internal eval set, if you have one, against both Claude Code on Opus 5.5 and Codex on Sol specifically, rather than relying on last month's comparison. Frontier model swaps inside an existing agent product change behavior in ways that do not always track a single aggregate score, and the tasks your team actually runs are the only ground truth that matters for a standardization decision.
Do watch for GPT-6 Luna's access terms. If Luna becomes available at a tier your organization qualifies for, the comparison worth running is Claude Code on Opus 5.5 against Codex on Luna, not Sol, because Sol was explicitly positioned as the more broadly available, not the most capable, half of OpenAI's split release.
When this applies to your stack
If your engineering org is running one coding agent as the default across most developers, and you have not re benchmarked since before September 22, this week's releases are a reasonable trigger to do it now rather than at the next scheduled review. The index gap is real, but it describes an average across a task mix that may not match yours, and the per task pricing difference is large enough to change the calculus for teams running agents at volume.
Contra Collective helps engineering organizations evaluate and roll out agentic coding tooling against their own codebases and workflows rather than off a leaderboard alone. If you are weighing a Claude Code or Codex standardization decision this quarter, that is a conversation worth having before you commit a team to either default.
FAQ
Is Claude Opus 5.5 the same thing as Claude Code? No. Claude Opus 5.5 is the underlying model. Claude Code is Anthropic's coding agent product, which uses Opus 5.5 as its default model as of September 22, 2026. The Coding Agent Index scores the combined product, not the model alone.
What is the difference between GPT-6 Sol and GPT-6 Luna? Both shipped September 22, 2026, as part of a split release. Sol is the generally available tier that Codex uses by default. Luna sits behind a more restricted, security focused access tier that OpenAI has not opened to general developer use as of this writing.
Does a higher Coding Agent Index score mean fewer bugs in production? Not directly. The index measures performance against a fixed task battery under controlled conditions. It correlates with general capability but is not a substitute for evaluating an agent against your own codebase, review process, and task mix.
Will this benchmark gap hold up over time? Treat it as a snapshot. Both models are two days old as of the September 24 refresh, and OpenAI has already signaled a more capable, gated tier in Luna that could change the comparison once it is broadly available.
More from the lab.
vLLM vs SGLang: Choosing a Production LLM Inference Engine in Late 2026
vLLM and SGLang are both open source, both built to serve LLMs at production scale, and both frequently recommended for the exact same job, which is exactly why teams keep picking the wrong one. The real difference is not a benchmark number. It is which attention and caching architecture matches your actual traffic pattern, and most teams never check that before they commit infrastructure budget to one or the other.
Ollama's MLX Backend Beats Metal by 93 Percent: Apple Silicon Local Inference Tested (October 2026)
Ollama's 0.19 release notes say it plainly: on Apple Silicon, Ollama is now built on top of Apple's MLX framework to take advantage of its unified memory architecture, rather than the Metal backed llama.cpp path it shipped for years. On an M4 Max running Qwen3.5, that swap took decode speed from 57.8 tokens per second to 111.4, a 93 percent improvement, and prefill from 1,154 to 1,810 tokens per second. The catch is in the fine print: the MLX runner is in preview, it needs 32GB or more of unified memory to activate at all, and as of this release it only accelerates one model family. Here is what actually changed, what it costs to turn on, and where it still falls back silently to the slower path.
DGX Spark vs Mac Studio M5 Ultra: Local LLM Inference Benchmarked (October 2026)
The M5 Ultra Mac Studio shipping in September 2026 is a quad die chip with 1.2 terabytes per second of memory bandwidth and up to 256GB of unified memory, with a 512GB configuration arriving in late October. Against a single NVIDIA DGX Spark, it is not close: on a 27B dense model, the Mac decodes at roughly four times the speed. Cluster two DGX Spark units together over NVIDIA's GX10 networking instead, and the comparison flips for a different workload entirely, concurrent multi user serving. Neither machine is the right answer for every workload, and the decision comes down to whether you are serving one user or several at once.