GPT-5 Mini Launches with Native Web Browsing at $0.10/M Tokens
OpenAI released GPT-5 Mini, a cost-optimized model with native real-time web browsing and multimodal input support priced at $0.10 per million input tokens. The release targets high-volume production workloads where up-to-date information retrieval is a core requirement.
Original sourceOpenAI launched GPT-5 Mini today, a smaller, cheaper variant of its GPT-5 model line designed for production deployments that need to process large volumes of queries without sacrificing access to current information. At $0.10 per million input tokens, it sits significantly below the pricing of full GPT-5 and positions itself as the default choice for teams building pipelines that previously had to choose between cost and freshness.
The headline capability is native real-time web browsing baked directly into the model rather than offered as a separate plugin or tool-call layer. This means developers don't need to wire up a separate search API, manage retrieval pipelines, or handle the orchestration overhead that previously came with building browse-augmented applications. Multimodal input support ships alongside browsing, accepting images and text in the same request.
The $0.10 per million input token price point is notable context: it undercuts several mid-tier retrieval-augmented generation setups when you factor in the saved cost of external search APIs and the engineering time to maintain them. For teams currently running Bing Search or similar alongside a cheaper model, the consolidated pricing may actually pencil out favorably even before accounting for latency improvements.
What remains to be seen is throughput behavior under load, rate limit structure, and how OpenAI handles the editorial and reliability questions that come with grounding model outputs in live web content. Real-time web access introduces freshness but also surfaces misinformation, paywalled content, and pages that actively try to manipulate crawlers — all problems the product documentation will need to address clearly.
Panel Takes
The Builder
Developer Perspective
“The DX bet here is collapsing the browse-augment stack into a single API call, and that's the right call — maintaining a separate retrieval pipeline with its own auth, rate limits, and error surface is a real cost most tutorials skip over. The question I'd actually want answered before shipping this in production: what does the tool-call trace look like, can I inspect which URLs were fetched, and is there a way to scope or restrict the browsing to specific domains? If the answer to all three is 'no, it's a black box,' that's a skip regardless of price.”
The Skeptic
Reality Check
“Native browsing in a mini model sounds great until you remember that 'real-time web' means the model is now grounding answers in whatever it can scrape, including SEO spam, outdated cached pages, and sites specifically optimized to manipulate LLM crawlers. The $0.10 price kills the RAG-pipeline cost argument, but the business case depends entirely on whether OpenAI publishes actual accuracy and hallucination rates for browse-augmented outputs — which they have not done here. What kills this in 12 months isn't a competitor, it's enterprises discovering their customer-facing chatbot confidently cited a Reddit thread from three weeks ago.”
The Founder
Business & Market
“The pricing architecture is genuinely smart: at $0.10 per million input tokens with browsing included, OpenAI is pricing out the middleware layer — the Perplexity-style retrieval wrappers, the search-API aggregators, the 'AI search' startups that were reselling Bing at a markup. The buyer here is the engineering team that was previously paying separately for model inference and search, and this consolidates both line items into one vendor relationship, which is also how OpenAI builds the lock-in. The moat isn't the model, it's the distribution: if this gets into enough production pipelines before a competitor price-matches, switching cost is the product.”
The Futurist
Big Picture
“The thesis GPT-5 Mini is betting on: within two years, the default application architecture for information-retrieval products won't include a separate search API — the model itself is the retrieval layer. That's a falsifiable claim, and this launch is a direct test of it. The second-order effect that nobody is talking about is what this does to the search advertising model: if production apps route information queries through a model instead of a search engine, the click never happens and the ad impression never registers — GPT-5 Mini at scale is structurally a defection from the search ad economy, not just a product launch.”