OpenAI shipped more than 30 features and products at DevDay 2026 and cut API prices for GPT-6.1 Sol, yet developers walked away unimpressed. The reason is simple arithmetic. Per-token prices fell, but the $200 Pro subscription now buys less usage, and the ultra-fast model, Astra Ultra-Fast, sits behind a $500 monthly plan. For anyone running coding agents, the real question is no longer the price per million tokens. It is how much a finished task costs, and where that money actually goes.
Why a big launch landed flat
Expectations were the first problem. Anthropic’s release of Claude Opus 5.5 had raised the bar for developer tooling, and the DevDay lineup did not clear it.
The second problem was volume. With more than 30 announcements arriving at once, the most useful pieces got buried. Spaces, a shared workspace where people and AI agents edit Markdown files, spreadsheets and slides together, was arguably the real highlight. Improvements to MCP (Model Context Protocol, the standard that connects AI models to outside tools and data) also deserved more attention. Technical glitches during the live demos did not help.
The third problem was the subscription change. OpenAI leadership publicly explained that Pro subscribers on the $200 plan would get less API-equivalent usage than before. Announcing the cut openly, rather than quietly degrading the service, was the more honest choice. It still frustrated paying users.
Three pressures behind OpenAI’s pricing
The DevDay decisions make more sense as a response to three problems OpenAI had to solve at the same time.
| Pressure | What happened | Effect on developers |
|---|---|---|
| Price competition | GPT-6.1 Sol launched far below its planned price to compete with Claude Opus 5.5 | Cheaper API tokens |
| Subscription subsidies | The gap between $200 plans and real API costs had to shrink | Smaller usage allowances |
| Ultra-fast serving costs | Astra Ultra-Fast reserves roughly six GPUs per request | A $500 plan and quotas that drain quickly |
Price competition
GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens. Its prompt cache read price, the charge for re-reading input the model has already processed, is $0.10 per million tokens. OpenAI widened the usual 90% cache discount to 95%, the cheapest cache rate yet seen in this capability tier. The larger GPT-6 Astra was designed to sell at $125 per million output tokens but launched at $50, cutting gross profit per output token by more than half. Lower prices pulled in so much traffic that OpenAI ran short of compute and temporarily paused new sign-ups for its $200 plan.
Subscription subsidies
Data from SemiAnalysis shows how differently the two leading labs subsidize their top consumer plans.
| $200 monthly plan | API-equivalent value per month | Multiple of the subscription price |
|---|---|---|
| OpenAI ChatGPT | About $2,084 | About 10x |
| Anthropic Claude | About $11,726 | About 58x |
Anthropic’s level of subsidy looks hard to sustain, and a pricing correction may come as it prepares to go public. For now, though, a Claude subscription buys far more compute for the same $200.
Spending data from Andreessen Horowitz explains why labs care so much about this small group. Among paying AI users, the top 1% spend an average of $1,000 a month, the top 10% spend $150 and the top 40% spend $27. Only about 2.3% of US households pay for an AI product at all.
The cost of speed
Astra Ultra-Fast removes queueing by dedicating hardware to each request, which makes every token expensive to serve. Testers received a $1,000 credit before launch, and two pull request reviews used up $650 of it. On the $20 ChatGPT tier, a few quick tasks on Astra at a low reasoning setting consumed 90% of a weekly quota in under 30 minutes.
Cache reads drive the bill
Coding agents re-read the same context on every step, so their cost profile looks nothing like a chat session. Across frontier models, cache reads typically make up 60% to 75% of total spending on agent projects. A seven-day breakdown of about $3,500 spent on Astra shows the pattern.
| Cost item | Amount |
|---|---|
| Regular input tokens | $770 |
| Regular output tokens | $412 |
| Cache reads | $2,400 |
That makes the cache read rate the number to compare first. Astra charges $1.00 per million cached tokens, while GPT-6.1 Sol charges $0.10, a tenfold gap. For context-heavy agent work, moving from Astra to GPT-6.1 Sol can cut total costs by 60% to 70%. Claude Opus 5.5 and Sonnet 5.5 both charge $0.20 per million cached tokens.
▲ Cache reads driving agent costs
Long context adds a second trap. Once an Astra session passes roughly 272,000 tokens, its cache read price doubles to $2 per million tokens. At 800,000 tokens of active context, a single tool call to read a file or search code costs about $2. Combining the ultra-fast mode with a long context pushes output to around $450 per million tokens. A plausible explanation is memory: a context above 200,000 tokens can claim a large share of a GPU cluster’s high-bandwidth memory and block other requests.
Measure cost per task, not cost per token
OpenAI’s own message points the same way. Subscription prices are disconnected from compute costs, and a cheaper model can still cost more per task if it needs extra reasoning steps or produces bloated output. OpenAI argues that GPT-6.1 Sol needs fewer reasoning tokens to solve hard development tasks and wants developers to compare the cost of completed tasks rather than rate cards.
Real usage logs show why. One developer ran 4 billion tokens through Claude Opus 5.5 for $2,000 and 1.78 billion tokens through GPT-6.1 Sol for $400. Opus handled about twice the tokens but cost five times as much. On cost-versus-intelligence charts, GPT-6.1 Sol sits on its own, far cheaper than its peers. Comparing GPT-6.1 Sol at extra-high reasoning with Claude Opus 5.5 at medium, Opus costs about three times more and finishes tasks about 1.5 minutes faster.
Cheaper does not mean better, though. Claude Opus 5.5 still appears to be the stronger model for everyday coding, while GPT-6.1 Sol shines at deep investigations and tricky system configuration problems. The gap widens on multi-day autonomous work. Anthropic’s models tend to notice a wrong turn, backtrack and correct course. Astra tends to double down, so a planned 20-line pull request can grow into 2,000 lines of well-tested but useless code. On that kind of work, the model with the higher price per token may still be the cheaper one per result.
Pairing models from different labs
Developers trying to balance cost and quality increasingly combine models instead of picking one. One open-source coding agent tool recently merged an overhaul of about 400,000 lines that lets a running Claude session spawn sub-agents on other providers’ models to verify code and review changes. Before that, handing work to another model meant calling the Codex CLI separately or copying text between chat windows.
A typical review loop built on this setup works like this:
- Use Claude Opus to implement any change longer than five lines.
- Send the change to GPT-6.1 Sol through Codex for a critical review. Sol often catches edge cases and subtle bugs that Opus misses.
- Revise until both models agree, or until Opus rejects the critique with high confidence.
- Test and verify, then merge automatically only if confidence that nothing will break is above 8 out of 10.
Lightweight models have a clear role too. GLM 5.3 Flash and DeepSeek V4.1 Flash handle heavy, deterministic jobs such as triaging 1,300 pull requests, but they struggle at design decisions with no objective way to check the answer. Basic environment checks matter as well: one audit of a coding agent fleet found that 30% of all errors came from agents using zsh instead of bash.
▲ Two models trading code reviews
Assistants and workspaces still need work
OpenAI’s personal assistant agent, Dots, shows how far these products have to go. In one user’s experience it was useful about a third of the time, and an email alert it had been told to watch for got lost among 14 consecutive messages. The problem is the single-threaded design that piles every notification into one stream. Good human assistants manage attention instead, sending urgent items as one-line messages and routing less important ones to separate channels. AI assistants likely need the same kind of triage.
Discoverability is a related weakness. The ChatGPT desktop app changes often and scatters its features, and a Codex feature that renders interactive HTML graphs and pages inside responses existed for six months with few users noticing. A stricter rule, where every user-facing feature needs advance approval from a product owner, may be the better way to keep a tool coherent.
What to check before you pick a plan
DevDay 2026 brought cheaper tokens and less generous subscriptions at the same time. Before changing plans or models, check these points:
- Compare cost per task. Look at what finished tasks actually cost in your usage logs, not just the price per million tokens.
- Check your cache read share. Cache reads can make up 60% to 75% of agent spending, so compare cache read prices first.
- Watch long-context price tiers. Some models charge more past roughly 272,000 tokens, so keep context below that line where you can.
- Be careful with the $500 plan. Astra Ultra-Fast is hard to justify unless you already hit the limits of the $200 plan. If you need more usage, compare Claude subscriptions as well.
- Ask whether speed matters. Cutting a response from 60 seconds to 10 changes how you work. The difference between 2 and 10 minutes does not, because you switch tabs either way. If a task makes you look away, run it on a cheaper model.
- Lower your reasoning setting. Extra-high reasoning can overthink simple tasks. Use high for daily work and save extra-high for large refactors.