Gemini 3.7 Flash and GPT-5.6 Sol Ultrafast Split on Agent Access

Two speed plays launched on the same day

On August 13, 2026, Google and OpenAI both announced faster AI, but they gave access to it in very different ways. Google released Gemini 3.7 Flash as a generally available model with the ID gemini-3.7-flash. It shipped as a stable API model rather than an experimental preview1619. OpenAI announced Ultrafast, a new API service tier that runs GPT-5.6 Sol on Cerebras hardware at up to 750 output tokens per second. OpenAI says that is as much as 14 times faster than Standard processing7. Google's model went to every developer immediately. OpenAI's tier went only to a small group of invited customers, and it came with no published price12.

For teams building autonomous agents, the gap in access matters more than the speed numbers. Google is competing on price and broad availability. OpenAI is offering top-end responsiveness on its strongest model to selected customers. Looking back from October, the first approach has produced more working software for enterprises.

What Google shipped

Google called Gemini 3.7 Flash its "most intelligent workhorse model yet for coding and agents." It arrived just three weeks after Gemini 3.6 Flash12. Google's own figures show large gains on agent-style work. On DeepSWE v1.1, the model rose from 49.0% to 65.3%. On FrontierCode 1.1 Main, it went from 34.4% to 43.6%. On AutomationBench, which tests real business workflows, it climbed from 17.0% to 30.4%13. Google also says the model is more careful about multi-step planning and tool calls. The company presents this as fewer retries and less human oversight in engineering workflows13.

The price was set for high-volume agent work. Through December 31, 2026, the model costs $0.75 per million input tokens and $3.75 per million output tokens. Those rates double on January 1, 2027, to $1.50 and $7.5012. Coverage of the launch agreed that the discount is temporary. VentureBeat described the low-price period as a window to test whether fewer retries actually reduce total operating costs15. Emergent advised budgeting at the 2027 rate for any workload that will run into next year. It also noted that thinking tokens are billed at the output rate18.

Google made the model available through its main agent tools on day one: the Gemini API in AI Studio, the Antigravity agent coding environment, the Gemini Enterprise Agent Platform, and the Gemini Enterprise app13. Google's developer guide said 3.7 Flash would become the default model for Managed Agents in the Gemini API17. That service launched in May. It lets developers run autonomous agents that keep their state between steps inside isolated, Google-hosted Linux sandboxes30. Consumer access was narrower. The model went into Gemini Spark, Google's always-on personal agent, but only for AI Pro and Ultra subscribers12. Ars Technica noted that the regular Gemini chatbot stayed on 3.6 Flash14.

Independent speed measurements back up the "fast" label. BenchLM measures Gemini 3.7 Flash at 282 tokens per second, but with an 11.2-second wait before the first answer29. A reasoning-heavy Flash model can produce text quickly once it starts and still pause noticeably before it begins.

What OpenAI previewed, and what it held back

Ultrafast is a service tier, not a new model. It runs the same GPT-5.6 Sol that became generally available on July 9. Only the hardware changes2. OpenAI said its first testers came from coding, commerce, financial research, support, and other interactive uses. It framed the preview as a way to learn where an order-of-magnitude speedup creates the most value7. Named early customers include Jane Street, Podium, Basis, and Rogo1.

The practical difference is large. One analysis divides 750 by 14 and estimates that Standard Sol produces about 54 tokens per second. On that basis, a five-turn agent loop of 1,000 tokens per turn would take about 93 seconds on Standard and about 7 seconds on Ultrafast1. For agents a person is waiting on, such as voice assistants, incident response, or trading-desk research, that is the difference between waiting and working in real time.

Coverage pointed out the gaps just as consistently. OpenAI published no price, no general-availability date, and initially no separate way to select the tier through the API2. TechTimes stressed that every performance number came from OpenAI and Cerebras themselves, not from independent testers2. That includes Cerebras's claim of no quality loss10. Without a rate card, enterprise architects could not build a cost model. One analysis said teams should not build production systems on the tier until access and pricing are confirmed2.

Where the coverage diverges

Reporting on Google's release split on what the speed of releases means. Google and its partners framed the three-week turnaround as quick, feedback-driven improvement1218. Ars Technica read it more skeptically. It argued that Google may be keeping up an appearance of constant progress while Gemini 3.5 Pro, which Google promised for June, still has not shipped. It also pointed to reports that Gemini's coding skills have fallen behind rivals14. VentureBeat added that the launch came amid departures of senior AI staff. It cited Reuters reporting that internal disagreements and limited computing capacity had slowed releases15.

Both readings can be true, and for enterprise agent builders they lead to the same conclusion. Google is now putting its fastest improvements into the Flash tier. Toloka described this as a reversal: Flash is no longer the model you settle for when Pro costs too much26. Kingy's review put 3.7 Flash in the same middle band of capability as Claude Sonnet 5 and GPT-5.6 Terra, at a lower introductory price16. That does not make it a top-tier model. It does make it an inexpensive default for running agents at scale.

OpenAI's announcement drew a different split. ExplainX reported that the reaction divided between businesses that wanted access and paying subscribers who felt excluded. OpenAI had made no commitment to bring Ultrafast to Codex or ChatGPT6.

How it looks eight weeks later

Both launches now look like waypoints. Google released Gemini 3.8 Flash on September 2. It is pitched specifically for long-horizon software engineering, autonomous agents, and complex enterprise workflows, at the same introductory price2430. Google's managed Antigravity agent now runs on 3.8 Flash by default24. Google's cloud documentation lists a retirement date of January 28, 2027 for 3.7 Flash and tells customers to upgrade11. In practice, the model had a shelf life of about five and a half months. BenchLM now ranks 3.8 Flash as the fastest model scoring above 70 on its quality index, at 216 tokens per second29.

OpenAI moved on too. At the end of September it formally launched Ultrafast for its GPT-6 family. The trade-offs are very different from the Cerebras preview: up to 300 tokens per second, up to 6 times faster in the API and 8 times in Codex, at six times the Standard price22. OpenAI's documentation now says Ultrafast is broadly available for GPT-6 Astra and GPT-6.1 Sol. It still lists only preview access for GPT-5.6 Sol4. One analysis notes that getting 5.6 Sol on Ultrafast still means going through an OpenAI account team, with no self-serve signup and no public price3. Another points out that GPT-6.1 Sol is the only Ultrafast option with EU data residency5.

The outcome cuts against simple speed comparisons. The fastest product in the August announcements, Sol at 750 tokens per second, remains largely out of reach. The tier OpenAI made generally available is slower, and its cost is now known.

What it means for enterprise agents

Taken together, these launches suggest that for autonomous agents, availability and predictable cost matter more than peak speed. An agent that calls a model dozens of times per task needs a stable model ID, a known price, and the ability to run at scale. Gemini 3.7 Flash offered all three on day one. GPT-5.6 Sol Ultrafast offered none of them to the general market216.

OpenAI's experiment still matters. Ultrafast showed what agents that respond in real time could look like and attracted demanding early users in finance17. Its pricing has also shaped the market. Speed is now sold as a premium tier at a clear markup rather than built in as a default22.

The lesson for enterprise teams is about operations. Google is replacing Flash models in a matter of weeks and setting retirement dates within months1118, so regression testing and model-switching layers are basic requirements, not extras. On OpenAI's side, the tiers carrying "fast" or "ultra" in their names differ widely in speed and price1, and teams should check which one they are actually buying. Agent frameworks that treat the model as a swappable part are better placed to keep up with both vendors.