GPT-5.3 Codex Spark vs GPT-5.6 Sol: OpenAI Retires the Speed Play

GPT-5.3 Codex Spark is being retired. On September 11, 2026, Open​AI's core product lead, Thibault "Tibo" Sottiaux, said on X that the fastest Codex variant would be gone "next week" — a little over seven months after it launched — because usage had been steadily declining and better models are now available. That is an executive statement relayed through a social post, not a dated deprecation notice: Open​AI has not published a specific sunset date on its own deprecation documentation, and the only date on the record is "next week," which points at the week of September 14. What is not in doubt is the direction. For seven months GPT-5.3 Codex Spark was the most interesting open question in Open​AI's lineup — a model that generated over 1,000 tokens per second and still never earned a public price. The retirement answers that question with a no, and it leaves GPT-5.6 Sol, the $5 / $30 flagship released July 9, 2026, as the model that actually carries Open​AI's low-latency coding story. The comparison this page has always drawn — a speed preview against the flagship that arrived after it — now has an ending.

What Open​AI announced, and what it did not

The announcement itself is thin, and it is worth being precise about how thin. Sottiaux framed the shutdown as the model having "had a good run" and as time to "make room for the future," with declining usage and significantly better available models as the reasons. No replacement was named. No migration path was offered. No date was given beyond "next week." The retirement has been picked up by multiple independent outlets — BlockBeats, TechFlow, TokenPost, Cointelegraph's Chinese edition, and Digit among them — which is what makes it corroborated rather than a single post, but every one of them is relaying the same executive statement. Nobody has an Open​AI document to point at yet.

That matters for one practical reason: if you are planning around a deadline, the deadline is soft. "Next week" from September 11 means the week beginning September 14, and Open​AI has left itself room to move it. The safe reading is that GPT-5.3 Codex Spark is not something to build anything new on as of this week, and that any workflow depending on it should be migrated before mid-September rather than after.

What Spark was, and why the retirement is not a surprise

GPT-5.3 Codex Spark launched in February 2026 — reports give February 12 — as a research preview rather than a product. It was Open​AI's first model designed specifically for real-time coding, where the milliseconds between keystroke and response matter more than a benchmark point, and it was the first public milestone in Open​AI's January 2026 partnership with chipmaker Cerebras. It ran on Cerebras' Wafer Scale Engine 3, the dinner-plate-sized chip with four trillion transistors, which keeps model weights on-chip and sidesteps the memory-bandwidth bottleneck that dominates latency in ordinary GPU inference. Alongside the silicon, Open​AI cut per-round-trip overhead by about 80%, per-token overhead by about 30%, and time-to-first-token by about 50% through persistent WebSocket connections and Responses API changes, with the WebSocket path enabled by default for Spark.

The trade-offs were baked in from day one. Spark is a 128K-context, text-only model, which puts it far behind the current flagships on capacity. And on the agentic software-engineering benchmarks Open​AI itself cites, it lands well below the full GPT-5.3 Codex — 58.4% versus 77.3% on Terminal-Bench 2.0, both vendor-reported — because it defaults to a lightweight working style that makes minimal, targeted edits and does not run tests unless you explicitly ask it to. Speed was the product, and the product was a preview. A preview that is never promoted is a preview that eventually gets switched off, and seven months with no general-availability announcement was the signal long before this week's post.

The API question is now closed permanently

The headline number attached to Spark in third-party trackers is $1.75 per million input tokens and $14.00 per million output — identical to the full GPT-5.3 Codex. Treat that number as fiction, and now treat it as permanently fictional. Spark was offered to ChatGPT Pro subscribers through the Codex app, the Codex CLI, and the VS Code extension, with API access extended only to select design partners and its own separate rate limits rather than standard usage buckets. Open​AI never published a public API price for it, and with the model being retired, it never will. A price on a tracker for a model with no public endpoint was always a number you could not buy at; it is now a number for a model that will not exist.

The flip side is that the retirement is unusually clean. There are no API keys to rotate, no endpoint to repoint, no pricing table to re-plan against, because there was never a general API in the first place. The people actually affected are ChatGPT Pro subscribers who use Spark inside the Codex app and CLI, and their migration is a model-picker change, not an engineering project.

Where the real-time coding story actually went

The successor to Spark is not another small model. It is a serving tier on a large one. In August 2026, Cerebras and Open​AI announced "Ultrafast mode," a new tier in the Open​AI API that runs GPT-5.6 Sol at up to 750 output tokens per second — up to 14 times faster than Standard processing, on the same weights, with the intelligence of the Standard model unchanged. It entered limited preview on August 13, 2026 with a selected group of API customers, no published pricing, and access expanding as capacity allows. Cerebras says early testers include Jane Street, Podium, Basis and Rogo, and that on Humanity's Last Exam the Ultrafast tier answered the full 2,500-question set in about 11 hours against more than 78 hours for a comparable slower configuration. Those are vendor and partner figures, unreproduced, and the tier is still a limited preview with no price — so the honest status is "announced and partially available," not "buyable."

The strategic read is the interesting part. When Open​AI wanted sub-second coding in February, it built a small dedicated model and served it on Cerebras. When it wanted the same latency six months later, it put the frontier model on Cerebras instead. That is a different bet: rather than maintaining a fast, narrow model alongside a smart, slow one, Open​AI is collapsing the two onto one set of weights and differentiating on serving infrastructure. If that holds, there is no Spark-shaped thing to wait for — the fast version of every future flagship is a tier, not a model name.

What you can actually call today

Strip away the preview and the retirement, and the callable Open​AI coding lineup is two models with published prices.

• GPT-5.3 Codex — $1.75 per million input tokens, $14.00 per million output, cached input at $0.175. A 400K-token context window with a 128K output ceiling. Open​AI's agentic coding model, in the public API since February 2026.

• GPT-5.6 Sol — $5.00 per million input and $30.00 per million output on requests up to 32K input tokens, rising to $10 / $45 above that, with cached input at $0.50 and cache writes at $6.25. Roughly a 1.05M-token context with a 128K output ceiling and text, image, and file input. Open​AI's flagship, released July 9, 2026.

Both sit on OrcaRouter at the vendors' list prices with 0% markup — the $1.75 / $14 line and the $5 / $30 line are the prices here, not a reseller's version of them. A provider price cut is live on this side the same day the vendor makes it, because OrcaRouter passes the list price through instead of re-marking it.

GPT-5.3 Codex Spark vs GPT-5.6 Sol, dimension by dimension

• Speed — GPT-5.3 Codex Spark at 1,000+ tokens per second on Cerebras hardware, vendor-reported, the entire reason it existed; GPT-5.6 Sol at about 244 output tokens per second in our own measurements on OrcaRouter in August 2026. Not close on the base tier — though the Ultrafast tier closes it to roughly a third of the gap, for however many customers get access.

• Price — GPT-5.3 Codex Spark never had a public API price and now never will; GPT-5.6 Sol at $5 / $30 on the base tier, $10 / $45 above 32K input tokens.

• Context — GPT-5.3 Codex Spark at 128K, text-only; GPT-5.6 Sol at about 1.05M with text, image, and file input.

• Agentic coding — GPT-5.3 Codex Spark at 58.4% on Terminal-Bench 2.0, vendor-reported; GPT-5.6 Sol at 47 on Artificial Analysis's Intelligence Index v4.3, independently measured, against about $1.99 of compute per Index task. The Index is a composite of ten evaluations including Terminal-Bench v4.0, so it is not directly comparable to a single-benchmark score — but it is the only number here nobody at Open​AI wrote.

• Working style — GPT-5.3 Codex Spark made minimal edits and skipped test runs by default; GPT-5.6 Sol is the full agentic flagship with reasoning and tool use.

• Access — GPT-5.3 Codex Spark was a ChatGPT Pro research preview, retiring the week of September 14, 2026 per an Open​AI executive statement; GPT-5.6 Sol is a public API model with a published price.

If the only thing you want is a fast first draft of a small edit, Spark's design was ideal — and as of this week that option is closing. If you want an agent that reads a repository, plans a change, runs the tests, and iterates until it passes, Spark was never trying to be that, and GPT-5.6 Sol is the model that is.

If Spark was in your workflow, this is the migration

For most readers the honest answer is that nothing changes, because GPT-5.3 Codex Spark was never callable outside a Pro subscription. If you were one of the Pro users running it in the Codex app or CLI, the replacement decision is straightforward: GPT-5.3 Codex for coding-shaped agent work at $1.75 / $14, and GPT-5.6 Sol for the hard, long-context, multimodal calls at $5 / $30. Both have published prices, both are in the public API, and both are reachable through one OrcaRouter endpoint on one key, so the switch is a model-string change rather than a second vendor contract. If a workload turns out to be sensitive to latency once Spark is gone, the routing DSL is what makes the fallback survivable — send the short interactive edits to GPT-5.3 Codex and escalate the long-horizon work to GPT-5.6 Sol, with automatic failover if a provider rate-limits you mid-run.

The one thing not to do is wait for a Spark successor. There is no announced one, and the evidence from August points the other way: Open​AI's answer to fast coding is a Cerebras-served tier on the flagship, currently in limited preview with no published price. That is worth watching and impossible to plan on.

What this retirement actually tells you

Three things are worth carrying forward. First, a "research preview" that runs seven months without a general-availability announcement is a product decision in slow motion — the retirement notice is the end of a process, not a surprise. Second, the pricing signal was available the whole time: a model that never got a public price never got the internal green light to be one. Third, the latency race has moved up-market. The interesting question is no longer whether a small model can be fast, but whether a frontier model can be fast enough to matter — and that is now a serving-infrastructure question, which is why Cerebras keeps appearing in Open​AI announcements and why a chip partnership reported at more than $20 billion is the thing to watch rather than the model name.

GPT-5.3 Codex Spark earned its place: it was the first real answer to "what if latency is the feature," and 1,000 tokens per second on a wafer-scale chip was genuinely new. It also never became something you could buy. The models you can buy are GPT-5.3 Codex at $1.75 / $14 and GPT-5.6 Sol at $5 / $30, both on one endpoint at list price with 0% markup and automatic failover — and if Spark's retirement has a practical lesson, it is that the callable model is the one worth building on.