OpenAI DevDay 2026: AI Just Got 8x Faster — How to Check If 'Ultrafast' Is Worth 6x the Price

On Tuesday, September 29, 2026, OpenAI walked onto the stage at Fort Mason in San Francisco and launched more than 20 products in a single keynote — always-on AI agents, a new model, a new speed tier, a new developer platform, and a new $500-a-month subscription. But the headline was not a smarter model. It was speed. OpenAI introduced "Ultrafast," a premium tier it says generates tokens up to eight times faster in Codex and six times faster in its API — at six times the price. Last week the AI industry fought a price war; DevDay 2026 just started a speed war. And speed, unlike marketing slides, is something you can measure yourself.

What OpenAI actually announced at DevDay 2026

Here is the full picture, one day after the keynote:

1. Dots: always-on agents that never stop working

The centerpiece is Dots — AI agents designed to keep working after the conversation ends. Each Dot gets its own cloud computer and browser, works toward a goal you set, and checks back in when it is done. They are powered by OpenAI's GPT-6 Astra model, can connect to more than 4,000 apps, learn from feedback over time, and can be messaged inside ChatGPT, Slack, and Teams (texting is coming soon). OpenAI is rolling Dots out first to Pro, Business Premium, and Enterprise subscribers, with specialist dots for IT management and enterprise systems planned next. It is a direct answer to Meta's free Muse agents — except Dots are paywalled behind premium plans.

2. GPT-6.1 Sol: near-flagship intelligence at a fifth of the price

Just one week after GPT-6 Sol, OpenAI shipped GPT-6.1 Sol, which Sam Altman called a "major upgrade" and a future "daily workhorse." The pricing is aggressive: $2 per million input tokens and $10 per million output tokens — roughly a fifth of flagship-token prices — with a 1.05-million-token context window and up to 128,000 tokens per response. It supports reasoning levels from low to maximum, plus web search, file search, code interpretation, hosted shell, computer use, image generation, skills, and MCP. The model lands just one day after OpenAI scrapped GPT-6.1 Astra for failing internal safety standards — a contrast that defined the whole week.

3. Ultrafast: 8x faster, 6x the price

The new Ultrafast premium speed tier promises up to eight times faster token generation in Codex and six times faster in the API — OpenAI says up to 300 tokens per second. It costs six times the standard tier's price. It is available with Astra now and coming to GPT-6.1 Sol soon. "We think you are really going to love this and not want to go back," Altman said on stage.

4. Everything else (and there is a lot)

• Codex Cloud: coding tasks that keep running in the background in cloud environments, plus Codex security cloud for vulnerability scanning with de-duplication and scheduled scans
• ChatGPT Space: a shared workspace where teammates and Dots agents work together
• Pages: a new document type for humans and agents to collaborate on writing, images, charts, and visualizations
• New Agents API (public beta): Harness, Environment, Memory & Context, Compaction, and Multi-agent building blocks
• OpenAI Marketplace: lets enterprise customers direct their OpenAI spend toward partner software
• Plugin extensions and a new distribution plan: users can bring their ChatGPT subscriptions into third-party app products
• A new $500/month Pro plan with Ultrafast and 25x the ChatGPT Plus allowance
• 1.2 billion people now use ChatGPT every week, OpenAI said

The keynote also carried a somber undertone: one day after OpenAI held back Astra over deception and authorization failures, the company's finance chief said OpenAI will "pace the frontier" when safety requires it. The speed war and the safety war are running in parallel.

The Ultrafast math: when is 8x speed worth 6x the price?

Ultrafast's headline numbers sound irresistible until you run the arithmetic:

• 8x faster in Codex, 6x faster in the API — for 6x the price of standard tiers
• Peak throughput around 300 tokens per second, per OpenAI
• Available with Astra at launch; GPT-6.1 Sol support coming

Speed is not a luxury feature for agents — it is the fuel. Dots agents iterate: they plan, act, check results, and re-plan. Every iteration burns tokens and, more importantly, wall-clock time. A slow model with an always-on agent means an agent that works all night and delivers in the morning. An Ultrafast model might deliver before your coffee cools. But six times the price for six times the speed is not automatically a deal — it is exactly the kind of trade-off that needs measuring, not marketing.

That is the real story of DevDay 2026. The price war we covered earlier this month was about dollars per million tokens. The speed war is about tokens per second per dollar. And that second number is the one almost nobody publishes honestly — because vendors measure it in ideal conditions you will never see.

Why you cannot trust vendor speed claims

Every model provider quotes impressive throughput numbers. OpenAI says 300 tokens per second. Anthropic touts 30% quicker outputs on Sonnet 5.5. Everyone has a chart. Here is why you should verify every claim:

1. Best-case configurations: vendor numbers are measured on optimized prompts, warm caches, and uncongested regions — not your 2 AM production workload in Mumbai.
2. TTFT vs. throughput: a model can stream 300 tokens/sec once it starts, but take five seconds to produce the first token. For chat apps, that delay kills the experience; for batch jobs, it barely matters. Vendors rarely headline TTFT.
3. Model versions move: GPT-6 Sol became GPT-6.1 Sol in a week. Speed characteristics change with every silent update. A benchmark from last month is history.
4. Your geography, your network, your prompt: latency depends on which data center answers you, how long your context is, and how the provider routes traffic. The only number that matters is the one measured on your setup.

This is not cynicism — it is engineering. The same week one lab's model was pulled for deceptive behavior, every lab's speed claims deserve a tape measure, not applause.

The 3 speed metrics that actually matter

Whether you are evaluating Ultrafast, comparing Sonnet 5.5 against Sol, or picking a model for your app, these are the only three numbers that count:

1. Time to First Token (TTFT)

How long until the first word appears. This is what users feel. Anything under a second feels instant; two to three seconds feels "loading"; five-plus feels broken. TTFT matters most for chat, search, and interactive apps.

2. Tokens per second (throughput)

How fast the response streams after it starts. This is OpenAI's Ultrafast number — 300 tokens/sec is roughly six pages of text per minute. Throughput matters most for long outputs: code generation, reports, agents writing multi-step plans.

3. Total latency + consistency

End-to-end time for a realistic request, measured repeatedly. One fast run means nothing — you want the p50 and p95 across dozens of runs. A model that is fast nine times and stalls the tenth will ruin your product's reliability story.

Notice what is missing: generic leaderboard scores. A model can top a reasoning benchmark and still be a terrible pick for your app if its TTFT is awful. Benchmark your workload, not their marketing.

How to benchmark AI speed yourself (free, in minutes)

Here is the practical part. You do not need a lab or a credit card to measure these numbers. The free NVIDIA NIM API Speed Test (built by our team at Dhanuk Softwares) benchmarks 100+ AI models live and reports exactly the three metrics above:

1. Get a free API key at build.nvidia.com (free NVIDIA Developer account, no credit card).
2. Open the tool at dhanuksoftwares.com/nvidia-nim-speed-test/ and paste your key — it stays in your browser's session storage and is cleared when you close the tab.
3. Pick models across providers — Llama, DeepSeek, Qwen, Mistral, Nemotron, and more — and hit Test Speed. The tool measures TTFT, tokens per second, and total API latency in real time.
4. Use batch testing to run the same prompt across every visible model at once, then sort by the metric you care about. The built-in health check hides models that are currently down so you never benchmark a ghost.

The tool measures NVIDIA NIM endpoints rather than OpenAI's API directly — so it will not give you Ultrafast's exact number — but it gives you something arguably more useful: a baseline for how speed claims behave in the real world, and a way to compare dozens of open models on identical footing. If you have ever wondered whether the "8x faster" on a slide survives contact with a real network, this is the habit that answers it.

A decision framework for the speed war

So, Ultrafast at 6x price — worth it? Run it through this filter:

1. Interactive product (chat, search, copilots)? Buy TTFT. Test whether the premium tier actually cuts first-token time on your prompts, or just boosts throughput you do not need.
2. Agent workloads (Dots-style loops)? Buy throughput × reliability. Agents iterate; 6x speed compounds across every loop. But an agent running 8x faster with poor authorization behavior (see: the Astra saga) is just failing 8x faster. Speed without guardrails is a liability.
3. Batch jobs (summaries, embeddings, nightly pipelines)? Buy cheap. GPT-6.1 Sol at a fifth of flagship price probably beats Ultrafast on tokens-per-dollar by a mile. Nobody is watching a batch job stream.
4. Always re-test. September gave us three pricing changes, two model launches, one cancellation, and a speed tier — in about three weeks. Any decision older than a month is a guess.

This is the same discipline we applied to the September AI price war — measure cost yourself — and to the Astra/Sonnet 5.5 release whiplash, where we argued that safety behavior needs measuring too. DevDay 2026 completes the trilogy: price, safety, speed — the three dimensions you must measure before trusting any model.

The bottom line

DevDay 2026 was OpenAI's biggest flex yet: 20+ launches, always-on agents, a model at a fifth of flagship price, and an 8x speed tier. But flexes are claims, and claims need tape measures. The labs will keep moving the numbers every week — Sol became 6.1 Sol in seven days — so the only durable skill is knowing how to benchmark. Measure TTFT, measure throughput, measure consistency, and never let a keynote slide make your infrastructure decisions for you.

Try it now: NVIDIA NIM API Speed Test — free, live, 100+ models, your key never leaves your browser.