Openai fires back at google with pocket-sized gpt-5.4 models that cost less than a cent

While Google crowed about Gemini 3 Flash’s speed crown, OpenAI spent the winter quietly rebuilding its smallest engines. On Tuesday the company uncorked two new ones—GPT-5.4 Mini and GPT-5.4 Nano—priced so low that running an entire customer-support bot for a week now costs less than a latte.

The summer gap that let rivals sprint

Since last June’s blockbuster GPT-5 drop, OpenAI’s road map revolved around brute scale: more parameters, bigger clusters, fatter invoices. That laser focus left the “tiny model” lane wide open. Google seized it, marketing Flash as the racer for developers who hate latency and invoices. The result: enterprise clients began benchmarking OpenAI’s flagship against Google’s bantamweight and asking awkward questions about value.

Nano and Mini are the belated answer. Both inherit the GPT-5.4 backbone but are distilled to 12 % and 3 % of the parent’s parameter count respectively, according to engineers familiar with the pruning schedule. The compression trade-off is real yet surprisingly gentle: Mini scores within 4 % of the full model on HumanEval coding tasks and beats GPT-4 Turbo on MMMU visual reasoning, sources tell Coastal Code.

Speed that turns cents into savings

Mini chews through 420 tokens per millisecond on an A100, roughly double GPT-5 Mini’s pace. That velocity lets start-ups stuff entire codebases into the 400 k-token context window, run diffs, and still finish before the coffee machine drips. Pricing lands at $0.75 per million input tokens and $4.50 output—a 70 % haircut versus the flagship.

Nano goes further into micro-budget territory: $0.20 in, $1.25 out. The catch? It’s API-only, no chat interface, and tuned for narrow chores—classification, column extraction, tiny code agents—where depth matters less than pennies. One beta tester routing 25 million monthly requests says the switch shaved $9,800 off last month’s bill with no measurable drop in intent-recognition accuracy.

Free-tier users get a silent upgrade

OpenAI slipped Mini into ChatGPT’s “Pensar” toggle overnight. Free and Plus subscribers who trigger the reasoning path now hit Mini instead of a throttled legacy model, explaining why some reported snappier answers before the press release even landed. Premium users will see Mini automatically sub in when the main engine is at capacity, a tactic that could cut peak-time GPU burn by up to 30 %.

Developers itching for Nano must wait until July for general API access, though the wait-list opened today. Early documentation shows a 128 k-token context—smaller than Mini but still large enough to swallow the average SaaS Terms-of-Service document whole.

The margin war begins now

OpenAI’s message is blunt: speed and thrift can coexist with frontier performance. Whether that is enough to lure shops already flirting with Gemini will hinge on real-world latency, not benchmarks. Google, meanwhile, is readying Flash-2 for autumn and has slashed internal cost targets by half. The titans aren’t just racing for IQ points anymore; they’re sprinting to the bottom of the price curve, and the cheapest brain may win the enterprise before the year is out.