Cutting your own price by 80% after three weeks means the first number was never real

Here's the deal: AI model prices going down is not news. They've been going down for three straight years. What carries information is the timing. Cut prices alongside a new generation and it reads as engineering progress. Cut them six months in and it reads as depreciation. Cut them twenty-two days after launch, by a factor of five, and only two explanations survive. Either the launch price had very little to do with cost, or somebody is breathing down your neck.

On July 30, 2026, OpenAI did exactly that. GPT-5.6 Luna went from $1.00 per million input tokens and $6.00 per million output tokens to $0.20 and $1.20 — a clean 80% off both sides. The mid-tier Terra went from $2.50/$15.00 to $2.00/$12.00, a 20% cut. The flagship Sol didn't budge from $5.00/$30.00. The GPT-5.6 family had shipped on July 8 (OpenAI's own post and Simon Willison's writeup date the public announcement July 9, a timezone artifact). So OpenAI shredded its own price sheet three weeks after printing it.

And most coverage skipped the part that actually explains the move. This wasn't only a price cut. The same day, OpenAI added a Fast mode for Sol: up to 2.5x the speed of Standard processing at 2x the Standard price, with, in OpenAI's words, "no change in intelligence." The docs confirm that Priority processing was renamed Fast mode on July 30, 2026, and that requests already tagged service_tier: "priority" keep working. So what actually happened on July 30 was not "prices came down." It was "the cheap end got cheaper and the expensive end got an upsell."

That asymmetry is the whole story. If serving costs had genuinely collapsed, the biggest, heaviest model should be the biggest beneficiary — kernel and batching wins scale with model size in absolute dollars. Instead Sol held firm and gained a premium tier. OpenAI didn't cut where cost fell. It cut where competition lives. And competition lives in exactly one neighborhood right now: the budget tier, where Chinese open-weight models and Google's Flash-Lite line have set up camp.

Who's standing around this price sheet

Start with OpenAI. GPT-5.6 isn't a model, it's three: Sol, Terra, Luna. All three share the same skeleton — a 1 million token context window, a 128,000 token maximum output, and a February 16, 2026 knowledge cutoff. The differences are size and price. At launch OpenAI claimed on its own benchmarks that Sol set a new high of 53.6 on Agents' Last Exam, beating Anthropic's Claude Fable 5 in adaptive reasoning mode by 13.1 points, and that Terra and Luna outperformed Fable 5 at roughly one-sixteenth the cost. All of that is vendor self-reported; treat it accordingly until third parties run it.

The stated reason for the cut is efficiency. Per InfoWorld's reporting, OpenAI attributed the decrease to serving efficiency improvements across the training and inference stack, and pointed at GPU kernel optimization work done for GPT-5.6 Sol that reduced inference cost without giving up performance. Notice the shape of that argument: efficiency discovered while building Sol is being cashed out as a discount on Luna. Nothing in the public materials explains why Sol's own price is unchanged if Sol is where the efficiency came from.

Second on stage: the Chinese labs. In the same week as the price cut, Alibaba shipped Qwen3.8-Max. It landed on QwenCloud on Monday, August 3 — a mixture-of-experts model with 2.4 trillion total parameters, 95 billion active, and a 1 million token context window. Alibaba's own benchmarks claim PaperBench 93.0 against Fable 5's 88.8 and IFBench 82.8 against 63.5, though Fable 5 still leads on SWE-bench Pro. Alibaba shares rallied 7% on the news. But the number that matters for pricing isn't in the benchmark table. It's the release schedule: open weights are due the following week, Alibaba's first open-source drop at this scale.

Open weights are a pricing event, not a research event. Once the weights are out, Alibaba stops setting the price of that model — whoever rents GPUs cheapest does. And 95 billion active parameters means the per-token serving cost is far below what a dense frontier model would need, because only those 95 billion actually fire on each forward pass regardless of the 2.4 trillion sitting in the file. A model with Terra-class serving economics and claimed Sol-class quality, distributed for free, squeezes OpenAI's middle tier from underneath. The Hacker News thread hit #2 on August 3 with 1,036 points and 556 comments, and the dominant sentiment wasn't awe — it was skepticism that only Alibaba's self-reported numbers exist with zero independent verification. That skepticism is fair and completely irrelevant to the pricing pressure, which lands either way.

Third: DeepSeek, which was already sitting below everyone. Its official price sheet lists DeepSeek-V4-Pro at $0.435 per million cache-miss input tokens and $0.87 output, with the lighter V4-Flash at $0.14 and $0.28. Cache hits are effectively free at $0.003625 and $0.0028. This is where the widespread "OpenAI undercut DeepSeek" summary falls apart. Luna's $0.20 input beats V4-Pro's $0.435 but loses to V4-Flash's $0.14, and Luna's $1.20 output is more expensive than V4-Pro's $0.87. OpenAI didn't go under the floor. It came down to the point where switching stops being worth the trouble.

Then Google and Anthropic. Google's official pricing lists Gemini 3.1 Flash-Lite at $0.25 input / $1.50 output, and the older 2.5 Flash-Lite at $0.10 / $0.40. Anthropic's docs put Claude Haiku 4.5 at $1 / $5, Claude Sonnet 5 at an introductory $2 / $10 through August 31 (going to $3 / $15 on September 1), and Claude Fable 5 at $10 / $50. Line those up next to the new OpenAI sheet and the targeting becomes obvious.

The numbers that changed, and the numbers they were aimed at

First, the cut itself, stated precisely. Luna dropped 80% on both sides: $1.00 to $0.20 on input, $6.00 to $1.20 on output. Headlines fixate on the input number, which creates a misleading impression, because most real workloads are output-dominated once reasoning tokens enter the picture. A model that thinks burns invisible tokens on the output meter. If only input had come down, real invoices would barely move. Because output fell by the same 80%, the savings are actually a full 5x.

Terra's 20% cut is a completely different animal. $2.50 to $2.00. Recognize that $2.00? It is exactly Claude Sonnet 5's introductory input price. That's not a number that falls out of a cost model; that's a number you read off a competitor's price page. And there's a clock on it: Sonnet 5's introductory rate expires August 31, with standard pricing of $3 / $15 taking effect September 1. So Terra at $2 / $12 looks like parity today and becomes 33% cheaper on input in four weeks. OpenAI staked out ground that gets more favorable while it sleeps.

Sol got no cut. It got Fast mode. Straight from OpenAI's developer account: "Fast mode for GPT-5.6 Sol in the API delivers up to 2.5x the speed of Standard processing at 2x the Standard price, with no change in intelligence." The documentation adds that both "fast" and "priority" work as service_tier values, and that Fast mode and Scale Tier get the same service-level agreement treatment, with service credits available on eligible Enterprise agreements when targets are missed. Actual latency targets are not published.

Item Before After July 30 Change
GPT-5.6 Luna input $1.00 / 1M tokens $0.20 -80%
GPT-5.6 Luna output $6.00 $1.20 -80%
GPT-5.6 Terra input $2.50 $2.00 -20%
GPT-5.6 Terra output $15.00 $12.00 -20%
GPT-5.6 Sol input / output $5.00 / $30.00 $5.00 / $30.00 unchanged
Sol Fast mode Priority Processing 2.5x speed at 2x price renamed and sped up
Ref — DeepSeek V4-Pro — $0.435 / $0.87 (cache miss) official price sheet
Ref — DeepSeek V4-Flash — $0.14 / $0.28 official price sheet
Ref — Gemini 3.1 Flash-Lite — $0.25 / $1.50 Google docs
Ref — Claude Haiku 4.5 — $1.00 / $5.00 Anthropic docs
Ref — Claude Sonnet 5 — $2 / $10 (through Aug 31) → $3 / $15 Anthropic docs
GPT-5.6 general availability July 8, 2026 cut on July 30 22 days

Translate that into something you can put on a slide. Say one agent turn consumes 20,000 input tokens and 2,000 output tokens. Under old Luna pricing: $0.020 plus $0.012, or $0.032 per turn. Under the new pricing: $0.004 plus $0.0024, or $0.0064. Exactly one-fifth. A product doing 10 million turns a month goes from $320,000 to $64,000. That isn't cost optimization — that's a different business case. That's the range where a proposal that got rejected last quarter gets approved this quarter.

And the best explanation for the timing is accounting, not engineering. Forbes reported that Uber burned through its entire 2026 enterprise AI budget within the first quarter after roughly 5,000 engineers started using Claude Code, and that Microsoft halted Claude Code licenses once spending blew past its annual budget within months. The same piece cites a Harness survey of about 700 leaders in which 29% of organizations said AI accounts for more than a quarter of total cloud spend, 42% review AI costs only quarterly despite week-to-week swings, and over 40% still track it in spreadsheets. When customers start flinching at the invoice, cutting your best-selling tier by 5x is not a defensive crouch. It's getting in front of the problem.

Who actually banks this

The clearest winners are teams running high-volume, low-complexity work: classification, field extraction, routing, moderation, summarization. These workloads have modest quality requirements and enormous call counts, which makes them brutally price-sensitive. Until now a lot of them died on the same sentence — "a model would be nicer, but rules are cheaper." At one-fifth the unit price that comparison stops working, because the salary of the engineer maintaining the rules engine now dwarfs the token bill by an order of magnitude.

Second: anyone building agents. Agent cost is not linear in step count. Every step re-sends the accumulated conversation as input, so context compounds, and a ten-step task doesn't cost 10x a one-step task — it costs 30x to 50x. That's why agents have historically demoed beautifully and incinerated budgets in production. InfoWorld quotes Chandrika Dutt of Avasant arguing that enterprise teams will spend the savings building more sophisticated agentic workflows that were previously hard to justify economically, and Pareekh Jain of Pareekh Consulting saying that for CIOs the biggest effect is scaling adoption rather than simply cutting cost — a Jevons paradox read.

Now the losers. First up, the "we'll cut your LLM bill by 70%" startups. Prompt compression, semantic caching, distillation-to-a-smaller-model — an entire category whose pitch was arbitraging the vendor's margin. You cannot sell a 70% saving against an 80% cut that the vendor just handed out for free. Several of those roadmaps became unfundable on July 30.

Second, and this is the sharpest edge of the whole move: self-hosting. At $0.20 per million tokens, running a modest open-weight model on your own GPUs is frequently more expensive than just calling Luna. People model self-hosting by looking at the GPU hourly rate, but the real bill includes idle capacity during off-peak hours, redundancy, model update cycles, evaluation infrastructure, and on-call staffing. An inference cluster running at 30% utilization almost always loses to an API. By pushing the budget tier this low, OpenAI is trying to kill the open-weight migration at the spreadsheet stage, before anyone gets far enough to benchmark quality.

Third, the party most likely eating this is OpenAI. Look at the shape of the price sheet and the intent is legible. Luna is now one twenty-fifth of Sol on input. At that spread there is only one rational architecture: put a router in front and send everything easy to Luna. OpenAI is actively encouraging that routing. Unit price drops, total token volume climbs, and the genuinely hard work still has to go to Sol, which did not get cheaper. Defend share at the bottom, harvest margin at the top. That arithmetic works only as long as Sol has no substitute — which is precisely the assumption Alibaba is aiming at next week.

Price wars that worked, and price wars that killed people

Start with a win. In July 2024 OpenAI launched GPT-4o mini at 15 cents per million input tokens and 60 cents output, describing it as an order of magnitude cheaper than previous frontier models and more than 60% below GPT-3.5 Turbo. That price sheet effectively invented the budget-tier category. A whole class of products that had never touched the API — bulk tagging, real-time chat, the unglamorous middle of data pipelines — showed up because the math finally worked. Cutting price to grow the market is a real strategy and it worked. The Luna cut is a restaging of that same script.

The second precedent is outside AI. On March 25, 2014, Google opened the cloud price war by cutting Compute Engine 32%, storage 68%, and BigQuery 85%. Amazon answered the very next day with an average 51% cut on S3. The consensus prediction was mutual destruction. What actually happened was that raw compute and storage genuinely commoditized while AWS kept its lead, because it stacked managed services on top of the layer it had just devalued. The lesson transfers directly: token prices trending toward zero does not decide the outcome, as long as you own the layers above — tool use, evals, observability, compliance, enterprise contracts.

Now the failures. The first is DRAM, and Korean readers know it in their bones. Through the 2007–2009 chicken game, leaders pushed prices below cash cost to force out weaker players. Germany's Qimonda filed for insolvency in January 2009 and Japan's Elpida entered bankruptcy protection in February 2012. The survivors won, but the industry torched years of profit getting there. AI inference rhymes with DRAM in two ways: fixed costs (GPU depreciation) are huge, and the less differentiated the product, the uglier the price competition gets. It also differs in one way that matters — DRAM was a fully standardized commodity and models are not, yet. How long that difference holds is the entire question.

The second failure mode is what happens when you take price all the way to zero without a business underneath it. Stability AI gave away image model weights, captured an enormous ecosystem, never converted it into durable revenue, and after its founder stepped down in March 2024 went through restructuring and rescue financing. Inflection AI raised $1.3 billion, had neither a cost advantage nor distribution, and was effectively absorbed into Microsoft as an acqui-hire in March 2024. Shared lesson: price is a weapon, not a business model. To win on price you need either a cost structure that survives the price, or a distribution channel that makes price irrelevant. One or the other.

And a reminder of how far a pricing shock can travel. After DeepSeek R1 landed, Nvidia lost $589 billion of market capitalization on January 27, 2025 — the largest single-day wipeout in US market history. The fear was that cheap models mean fewer GPUs. What followed instead was an explosion in token consumption. So reading this 80% cut as a bearish signal for infrastructure demand is exactly backwards, or at minimum premature.

How the other side punches back

Google has the easiest hand. Gemini 2.5 Flash-Lite already sits at $0.10 / $0.40, below Luna's new price, and 3.1 Flash-Lite is close at $0.25 / $1.50. More importantly Google serves inference on its own TPUs, which means it can match a price without worrying about anyone else's margin. Expect two moves: a repricing of the current Flash-Lite generation, and bundling that makes the price question moot — Gemini is already inside Workspace, Android, and Search, and "you don't have to buy it separately" beats any per-token number.

Anthropic almost certainly will not chase the bottom. Fable 5 at $10 / $50 is deliberate premium positioning, not an oversight. Its moat is reputation in coding and agentic workflows plus enterprise contracts, none of which are priced per token. But there's an awkward date on the calendar. Sonnet 5's introductory $2 / $10 expires August 31 and standard pricing of $3 / $15 kicks in September 1 — right after Terra settled at $2 / $12. Go through with the increase and Sonnet 5 becomes 50% more expensive on input than the equivalent OpenAI tier. Whether Anthropic holds that schedule over the next four weeks is the single best tell for how far this price war spreads.

Alibaba has already counter-punched, and its weapon is a different shape entirely. It isn't cutting price — it's removing price. Once Qwen3.8-Max weights ship next week, a pack of inference providers will start competing to serve it, and the floor gets set by GPU rental economics. OpenAI cannot price against that floor. What it can compete on is reliability, tooling, data governance, and procurement paperwork, which genuinely do decide enterprise deals. Alibaba has real weaknesses here too: as of August 3 there is no independent verification of its benchmark claims, and Chinese-origin models still hit friction in US and European procurement regardless of how good they are.

DeepSeek is already below everyone but shows signs of a scaling ceiling. Its official docs announce an upcoming peak/off-peak policy charging 2x during 9:00–12:00 and 14:00–18:00 Beijing time, with no confirmed effective date. That's not a company that wants more revenue per token — that's a company rationing capacity by shifting demand out of its busiest hours. However cheap the sticker, a tier that doubles in price or queues during business hours quietly drops off enterprise shortlists. It's plausible that OpenAI stopped at $0.20 rather than chasing $0.14 precisely because it priced that constraint in.

Finally, the resellers and the middleware. AWS Bedrock, Google Vertex, and Microsoft Foundry resell other people's models, so each has to decide whether to pass the cut through or absorb it as margin. And the routing layer — the middleware that decides which request goes to which model — is the quiet winner here. When the spread between your cheapest and most capable tier hits 25x, automated routing stops being a nice-to-have and becomes the highest-leverage line of code in your stack.

So what actually changes

For regular users, nothing changes directly. This is an API price sheet; ChatGPT subscription pricing is untouched. Indirectly, though, it matters a lot. The apps you already use — the summarize button in your notes app, the support chatbot on a shopping site, translation tools, inbox triage — mostly run on exactly this budget tier. At one-fifth the unit cost, expect more generous free tiers, lower paid tiers, or AI features appearing in products that couldn't previously afford them. Brace for the flip side too: the cost of generating garbage fell by the same 80%. Spam, fake reviews, and SEO sludge just got five times cheaper to manufacture.

For developers and practitioners, there's work to do this week. One: re-examine your routing. A lot of the traffic you send to Terra is there because of price, not quality, and the boundary just moved. But cheaper does not mean better — Luna is the same model it was on July 29. Run an actual A/B on your own data and measure quality regression before you migrate anything. Two: Fast mode. If service_tier: "priority" is hardcoded anywhere in your stack, you are on Fast mode right now. It still works, but you're riding a renamed product with different speed characteristics, so go look at a bill before you assume nothing changed. Three: instrument output tokens. On reasoning-heavy workloads, a lower unit price gets eaten alive if thinking tokens grow. Track cost per completed task, not cost per million tokens.

For investors, the question reduces to one thing: is this cost falling, or margin being spent to buy share? The audited evidence isn't public, but the shape of the price sheet argues for the second. A genuine cost breakthrough would show up first on the heaviest model, and three weeks is far too short a window for a new serving optimization to be validated and rolled into production. Meanwhile bolting a 2x Fast mode onto Sol is an explicit statement that OpenAI still has pricing power at the frontier. Put together, this looks like textbook price discrimination rather than a straightforward pass-through of lower costs. But there is no audited disclosure of OpenAI's actual inference margin, so anything beyond that is inference on my part, and I'll label it as such.

For policymakers, the interesting angle is different. The proximate trigger for this cut was price pressure from Chinese open-weight models, and that is not a pressure that export controls or tariffs can reach. Weights don't arrive on container ships. The moment Alibaba publishes a 2.4 trillion parameter checkpoint, that capability becomes reproducible in any datacenter on earth. For anyone writing a national AI strategy, the useful question stops being "where does our model rank" and becomes "whose weights are our companies building on." As price approaches zero, dependency stops being about cost and starts being about standards and data.

And if you own an AI budget, treat this as a chance to reallocate rather than to save. That's what both quoted analysts converge on, and it's what the Uber and Microsoft anecdotes demonstrate. When unit price drops, organizations do not spend less — they consume more. What you need right now is measurement, not a renegotiation. The scariest number in this whole story might be that over 40% of organizations are still tracking weekly-volatile AI spend in a spreadsheet they open once a quarter.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Directly, nothing. Your ChatGPT subscription didn't change; this is developer API pricing. Indirectly, the AI features in apps you already use just got 5x cheaper to run, so over the next few months expect bigger free tiers and new features. Just remember the cost of generating spam and fake reviews fell by exactly the same amount.

— Why now, of all moments? The three-week gap is the answer. In the same week, Alibaba shipped Qwen3.8-Max and announced open weights for the following week, and DeepSeek was already parked between $0.14 and $0.435 on input. Meanwhile enterprises were publicly blowing through AI budgets. This reads far less like technology getting cheaper and far more like a defensive line being pulled forward to protect the budget tier.

— Is OpenAI actually cheaper than the competition now? Depends on the tier. Luna's $0.20 beats DeepSeek V4-Pro at $0.435 but loses to V4-Flash at $0.14 and Gemini 2.5 Flash-Lite at $0.10, and Luna's $1.20 output is above V4-Pro's $0.87. Terra's $2 matches Sonnet 5's introductory price today and undercuts it from September 1. So the honest framing isn't "cheapest" — it's "cheap enough that switching isn't worth it." How long that holds is genuinely unknowable until Alibaba's weights are out and someone benchmarks them independently.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!