OpenAI GPT-5.4 Mini and Nano Bring Near-Flagship AI at Budget Prices

OpenAI's Small Models Get Serious

OpenAI has released GPT-5.4 mini and GPT-5.4 nano, which the company describes as its most capable small models yet, built to carry the flagship GPT-5.4's strengths into faster and far cheaper packages aimed at high-volume workloads1112. The mini model is the headline act: it improves on GPT-5 mini across coding, reasoning, multimodal understanding, and tool use while running more than twice as fast, and it approaches full GPT-5.4 on several evaluations1120. Nano, meanwhile, is positioned as the smallest and cheapest member of the family, recommended for classification, data extraction, ranking, and lightweight coding subagents1116.

The benchmark gap between mini and the flagship is remarkably thin. On SWE-Bench Pro, GPT-5.4 mini scores 54.4% against the flagship's 57.7%, and on OSWorld-Verified — a computer-use evaluation — it reaches 72.1% versus the full model's 75.0%, while demolishing GPT-5 mini's 42.0%1116. Nano posts 52.4% on SWE-Bench Pro and 39.0% on OSWorld-Verified16. On GPQA Diamond, a graduate-level reasoning test, mini scores 88.0% and nano 82.8%, against 93.0% for the flagship11. One clear weak spot: long-context retrieval drops sharply for the small models, with mini scoring just 33.6% on a 128K–256K needle-retrieval test compared with the flagship's 79.3%11.

The Pricing Story — and Its Fine Print

The affordability framing is real but has more fine print than the launch materials suggest. GPT-5.4 mini costs $0.75 per million input tokens and $4.50 per million output tokens, while nano runs $0.20 per million input and $1.25 per million output — figures consistently reported across coverage of the launch111417. Both support prompt caching, dropping cached input to $0.075 and $0.02 respectively, a 90% discount59. The full GPT-5.4, by comparison, runs $2.50/$15.00 per million tokens56.

But "affordable" is relative to what you were buying before. Third-party analyses note that mini's input price tripled from GPT-5 mini's $0.25, and nano's doubled from $0.05 — a point developers on Reddit flagged immediately, warning teams not to blindly swap model names in their configs29. Against the broader market, the picture is mixed: nano undercuts Gemini 3.1 Flash-Lite on both input and output pricing, but DeepSeek's open-weight models undercut nano on output-heavy generation, with DeepSeek V3.2 charging $0.42 per million output tokens against nano's $1.25910.

GPT-5.4 mini ships today in the API, Codex, and ChatGPT — supporting text and image input, tool use, function calling, web and file search, computer use, and a 400K-token context window — while nano is API-only for now1116. In Codex, mini consumes only about 30% of the GPT-5.4 quota, and Codex can delegate to it as a subagent so that less reasoning-intensive work runs on the cheaper model1118. In ChatGPT, mini is available to Free and Go users via the Thinking feature, and serves as a rate-limit fallback for GPT-5.4 Thinking for everyone else1120.

The Subagent Architecture Is the Real News

The most strategically significant thing about this launch isn't any single benchmark — it's the architecture OpenAI is explicitly endorsing. The company frames these models as purpose-built for a compose-and-delegate pattern: larger models plan and decide, while smaller models execute quickly at scale1112. Exchange4media's coverage of the release makes the same point, noting that OpenAI emphasized model selection is now driven by responsiveness and efficiency rather than raw size20.

This is where the competitive stakes get sharp, because OpenAI is not the only lab racing to the bottom of the price-performance curve.

Google Is Iterating Flash at a Frenzzy Pace

Google's Flash line is iterating faster than anything OpenAI ships. Gemini 3.6 Flash went generally available on July 21, 2026; Gemini 3.7 Flash followed on August 13; and Gemini 3.8 Flash superseded it on September 2 — all at identical $0.75/$3.75 per-million introductory pricing, all with one-million-token context windows3538. That context advantage matters: Gemini's Flash models accept text, image, video, audio, and PDF input and offer up to 65,536 output tokens, versus mini's 400K context and text-plus-image input38. Google has also been aggressive about retiring old models, with 3.6 Flash already scheduled for shutdown on November 19, 202639.

Anthropic's Answer Is Still Pending

Anthropic shipped Claude Opus 5.5 on September 22, 2026 and Claude Sonnet 5.5 six days later, with Sonnet generating output more than 30% faster than Sonnet 5 at unchanged $2/$10 per-million pricing2931. Anthropic has explicitly confirmed a third model, Claude Haiku 5.5, "built for high-volume and cost-sensitive applications" and due "in the coming weeks" — its direct answer to GPT-5.4 mini and nano's territory2931. Until it ships, Anthropic's budget tier remains Claude Haiku 4.5, which third-party comparisons place around $1.00/$5.00 per million — notably pricier than GPT-5.4 mini10. The Verge-adjacent coverage of the Sonnet launch reports testers describing it as a better collaborator than Sonnet 5, with speed suited to quick iteration31.

The Open-Weight Squeeze

The quietest but most consequential pressure on OpenAI's pricing comes from the open-weight ecosystem, which has spent 2026 closing the gap to the frontier at prices closed labs cannot match. DeepSeek's V4 family — V4-Pro at 1.6T parameters (49B active) and V4-Flash at 284B (13B active), both MIT-licensed with one-million-token context — put up 80.6% and 79.0% respectively on SWE-Bench Verified, numbers competitive with closed frontier models2123. DeepSeek V4-Flash API pricing has been reported as low as $0.054/$0.242 per million tokens on some providers24.

Open-weight releases have kept accelerating through the fall: GLM-5.3 and GLM-5.3-Flash from Z.ai landed in August, Kimi K3 from Moonshot in July at 2.8T total parameters, MiMo-V2.6-Pro from Xiaomi in September, and DeepSeek V4.1-Flash on September 10 — a 552B-parameter MoE activating only 8B to 16B per token262728. These models are not just cheap; several sit within a point or two of the leading closed models on agentic coding benchmarks27. The catch for enterprises is licensing: DeepSeek's MIT and Qwen's Apache 2.0 terms are genuinely unrestricted, while GLM-5.3, Kimi K3, and MiniMax M3 carry custom licenses with revenue thresholds and review requirements27.

Where This Leaves the Market

Reading across the coverage, the GPT-5.4 mini and nano launch is best understood not as a single product event but as one move in an escalating price-performance war across four fronts. OpenAI's pitch — near-flagship capability at 30% of flagship cost, with a native subagent architecture — is coherent and well-supported by the benchmark data119. But it comes with a real price increase over the previous small-model generation29, a long-context weakness the launch materials don't headline11, and an API-only nano that locks the cheapest tier out of ChatGPT and Codex entirely16.

The divergence in the coverage is telling. Gadget-adjacent outlets emphasize the affordability framing and broad accessibility1317, while developer-focused pricing guides flag the increases and the competitive undercutting from DeepSeek and Gemini910. SQ Magazine's take captures the split well: mini is "the real star," delivering near-flagship performance at a fraction of cost and time, while nano "quietly solves a big problem" for teams that don't need a powerful model for every task18.

The committed reading here: the multi-model, tiered-agent architecture is now the default way production AI gets built, and the winners of the next year will be the labs that make the cheap tier genuinely good. OpenAI's mini clears that bar convincingly; nano is competitive but not dominant; and Google's Flash cadence and the open-weight labs' pricing keep the pressure on everyone. When Claude Haiku 5.5 arrives, it will land in the most contested segment of the entire market — and on current evidence, it will have its work cut out.