The Stack

OpenAI retires its 1,000-tokens-per-second coding experiment

Editorial

Editorial

— 3 min read
Share
OpenAI retires its 1,000-tokens-per-second coding experiment

Two stories about what AI infrastructure is actually worth right now: one lab pruning a model that traded brains for speed, and an entire industry selling its power contracts to somebody else.

OpenAI is switching off GPT-5.3-Codex-Spark, the ultra-fast coding model it launched in February as a research preview, with retirement scheduled for the week beginning September 14. Thibault Sottiaux, who runs Codex, cited declining usage and better models now available; no replacement was named and no migration document was published. Spark was never a good model by the usual measures — it scored 58.4% on Terminal-Bench 2.0 against 77.3% for the full GPT-5.3-Codex, kept only a 128k text-only context window, and never got a public API price at all, which was the tell. What it did have was speed: more than 1,000 tokens per second on Cerebras' Wafer Scale Engine 3, making it the first production milestone of the compute deal OpenAI announced in January 2026, reportedly worth more than $20 billion across up to 750 MW of inference capacity.

The retirement matters less as a product death than as a change in where speed lives. OpenAI's answer to fast coding now is not a smaller model but an "Ultrafast" serving tier that runs the flagship GPT-5.6 Sol at up to 750 output tokens per second on the same weights — limited preview since August 13, no published price. The lesson for anyone building on previews: a model that never receives a public price is a model that never got internal permission to be a product.

Donald Trump's pledge to make Bitcoin mining an American industry is coming apart, and AI data centers are the ones taking the pieces. Bloomberg reports that miners are converting facilities to AI and high-performance computing as mining economics collapse — Bitcoin's total market value sits roughly $1 trillion below its October 2025 peak, while CoinShares puts the weighted average cash cost of producing one Bitcoin among listed miners at about $79,995 in the fourth quarter of 2025, the hardest quarter since the April 2024 halving. Miners including Hut 8, IREN and Core Scientific have signed more than $70 billion in AI-related contracts, some expect AI to supply 70% of revenue by the end of 2026, and the group's roughly 27 GW of connected power is the asset being repurposed. We looked at what happens when that build-out goes wrong — AI data center risk is heading for the catastrophe bond market.

The read: the scarce commodity was never the chip or the rig. It was a grid connection with a queue number, and whoever holds one can rent it to whoever pays best — crypto just happened to be the tenant that moved out.

What to watch: whether Cerebras' Ultrafast tier gets a public price before OpenAI's next flagship, since a tier nobody can buy is a preview with better branding.

If latency is now a serving tier instead of a model, does a lab still need a fast-and-cheap option at all? Tell us in the comments.

Read more

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon turned its always-listening Bee line into a two-device family on Thursday, and NVIDIA opened a 2.8-trillion-token Indic corpus recipe. The connecting thread: both are bets that the next AI money is in memory — yours, or your model's. Amazon put two new always-listening devices on sale, and both now lead to a subscription. Bee Frontier, the next-generation wristband that also clips to clothing, ships now at $74.99 in Onyx, Lavender and Sage — up from the $50 predecessor — with seven-day