GPT-5.3 Codex Spark vs GPT-5.6 Sol: OpenAI's Speed Play Isn't on the API (Yet)

GPT-5.3 Codex Spark is the fastest coding model ever shipped — a research preview that generates over 1,000 tokens per second on Cerebras silicon and finishes code edits up to 15x faster than the full GPT-5.3 Codex, a vendor claim. Here is the part most coverage buries: it is not a general API product. There is no public API price, access is limited to ChatGPT Pro subscribers through the Codex app and CLI, and it is not available on OrcaRouter. The two coding models you can actually call through one endpoint today are GPT-5.3 Codex at $1.75 per million input and $14.00 per million output, and GPT-5.6 Sol at $5.00 / $30.00 — and the comparison between the speed preview and the flagship is what tells you whether Spark is worth waiting for.

The timeline explains the confusion. GPT-5.3 Codex, OpenAI's agentic coding model, reached the API in February 2026; the Spark variant followed the same month as a research preview aimed squarely at interactive coding, where the milliseconds between keystroke and response matter more than a benchmark point. Five months later OpenAI shipped the GPT-5.6 family — Sol, Terra, and Luna — and GPT-5.6 Sol, the $5 / $30 flagship released July 9, 2026, is the model that made most of the speed gap academic. This is not a matchup between two rivals; it is a matchup between a research preview and the flagship that arrived after it.

What GPT-5.3 Codex Spark actually is

GPT-5.3 Codex Spark is OpenAI's first model designed for real-time coding, and the first public milestone in its January 2026 partnership with chipmaker Cerebras. It runs on Cerebras' Wafer Scale Engine 3, the dinner-plate-sized chip with four trillion transistors, which lets it sidestep the GPU-cluster communication bottleneck that dominates latency in ordinary inference. Alongside the chip, OpenAI 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 — the WebSocket path is enabled by default for Spark.

The trade-off is blunt. On the agentic software-engineering benchmarks OpenAI cites, Spark 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 automatically run tests unless you explicitly ask. It is a 128K-context, text-only model, which puts it far behind the current flagships on capacity. Speed is the product, and the product is a preview.

The pricing situation: no API price, and what the trackers say

OpenAI has not published an API price for GPT-5.3 Codex Spark. At launch it was available to ChatGPT Pro subscribers through the Codex app, the Codex CLI, and the VS Code extension, with API access offered only to select design partners, and it carries its own separate rate limits rather than counting against standard usage. Half a year later that is still the honest summary. A few third-party pricing trackers now list a $1.75 / $14.00 per-million line for Spark — identical to the full GPT-5.3 Codex — but OpenAI has made no general-availability announcement, and a number on a tracker is not a number you can buy at. Treat every Spark price as unconfirmed until OpenAI says otherwise.

What you can buy today, with published prices:

• GPT-5.3 Codex — $1.75 per million input, $14.00 per million output, cached input at $0.175. A 400K context window with a 128K output ceiling. OpenAI'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; cached input $0.50, cache write $6.25. A roughly 1.05M-token context with 128K output, text plus image plus file input. OpenAI'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 — Spark at 1,000+ tokens per second on Cerebras hardware (vendor-reported), the entire reason it exists; GPT-5.6 Sol's measured output on OrcaRouter over the last 7 days is about 244 tokens per second. Not close — and not the whole story.

• Price — Spark has no public API price; GPT-5.6 Sol at $5 / $30 on the base tier, $10 / $45 above 32K input tokens.

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

• Agentic coding — Spark scores 58.4% on Terminal-Bench 2.0 (vendor-reported); GPT-5.6 Sol posts an AA Coding score of 77.4, independently measured by Artificial Analysis at #2 of 132 models.

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

• Access — Spark is a ChatGPT Pro research preview; 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 is ideal. If you want an agent that reads a repository, plans a change, runs the tests, and iterates until it passes, Spark is not trying to be that — and Sol's 77.4 AA Coding score is the number to weigh against a preview with none.

The routing reality: you do not have to choose blind

This is one of those matchups where the routing layer changes the decision. Both callable models are on OrcaRouter at the vendors' own list prices with 0% markup — openai/gpt-5.3-codex at $1.75 / $14 and openai/gpt-5.6-sol at $5 / $30 — so you can run the same workload through both on one key and let your own traffic decide instead of trusting either vendor's narrative. The routing DSL makes that practical: a failover rule where the coding-shaped agent work runs on GPT-5.3 Codex and the hard, long-context calls escalate to GPT-5.6 Sol, measured per-task on your own data.

Three OrcaRouter specifics make this cheaper than the alternatives. First, $0 per-token markup — you pay the provider's list price, nothing on top. Second, BYOK — bring your own existing provider key and the vendor bills you directly, so there is no prepaid-credit pool and no router-inflated invoice to reconcile. Third, the endpoint is OpenAI-compatible: point your existing chat/completions or Responses client at the OrcaRouter base URL and the same code keeps working, with 200+ models behind a single endpoint for when you outgrow one vendor's lineup. The guardrails layer in front of that endpoint is the other reason a research-preview urge is safe to defer — until Spark has a public API and an SLA, the production answer is a published model, and the routing layer keeps that model at list price with automatic failover so a rate-limit spike does not stall a pipeline.

Where Spark is the wrong answer

Honest boundaries, because Spark's design is a compromise you should not stumble into. If your work is long-horizon agentic coding — a task that needs to understand a whole repository, hold a plan across many turns, and verify its own changes — Spark's minimal-edit, no-test-run default is a footgun, and its 128K text-only context will not hold the repository. If you need images or files in the prompt, it cannot take them. If you need a price you can forecast or an SLA you can rely on, it has neither. And if your workload is a high-throughput batch pipeline, 1,000 tokens per second does not matter; cost per token and throughput do, and those favor the cheaper GPT-5.3 Codex.

Also worth naming what we do not host: GPT-5.3 Codex Spark is not on OrcaRouter, because there is no public API to route to. We do not pretend otherwise, and the honest planning move is to build on the models that exist — GPT-5.3 Codex for coding-shaped agent work at $1.75 / $14, GPT-5.6 Sol for the hard, long-context, multimodal agent calls at $5 / $30 — and to treat Spark as a preview to watch, not a dependency to bet a pipeline on.

Bottom line

GPT-5.3 Codex Spark is the most interesting research preview OpenAI has shipped this year, and it is not a production decision yet. The production decision — the one with published prices, benchmarks you can verify, and an API that answers — is GPT-5.3 Codex for coding-shaped agent work and GPT-5.6 Sol for the hardest calls, both on one endpoint at list price with 0% markup. When Spark does reach a public API, it will almost certainly sit behind the same kind of endpoint, and switching will be a model-name change, not a migration. Until then it is a reason to watch — and the fastest way to be ready is to have the callable models wired up on a key you already own.