OpenAI And Anthropic Restart The Frontier AI Price War

Have you ever watched two people promise to take a breather, then sprint the second the other one blinks? That is roughly how Tuesday felt if you follow frontier models. One lab dropped a new flagship around lunch in New York. By early afternoon, the other had new names, new charts, and a price list that made the morning sticker look stale. I have covered product launches for years, and this one did not feel like a calm product cycle. It felt like a stare-down that spilled into the checkout line.

When A Pause Turns Into A Sprint

Ten days earlier, both chief executives had nodded along with the idea that the industry should ease off the accelerator. The phrase floating around was simple enough: pace the frontier. Investors heard prudence. Safety teams heard breathing room. Customers heard, maybe, fewer surprise upgrades landing in production on a Tuesday.

Then the calendar flipped. A new Opus-class model arrived with a pitch that sounded almost modest. Same neighborhood of quality as a much more expensive sibling, they said, for a lot less money. Fast mode if you want speed. Higher usage caps if you live inside the product all day. Cache reads cheap enough to matter if your agents keep rereading the same context. On paper, that is a customer-friendly update. In the market, it was a starting pistol.

Minutes later, the rival stack moved. Two mid-tier names hit the board with prices cut in half against the rate everyone had been staring at that morning. The workhorse landed at two dollars per million input tokens and ten dollars per million output. The lighter option went even lower, the kind of dime-and-fifty-cents structure that looks aimed at open-weight traffic rather than prestige buyers. Cached input on the workhorse dropped to twenty cents. The top-end model stayed expensive, which is a quiet way of saying the prestige SKU still needs a halo.

Higher usage limits and lower cost give you more flexibility and room to iterate.

That line is true. It is also incomplete. Flexibility is wonderful until the model behind a given call is not the one you thought you hired. I will get to that. First, the money, because money is what actually moved on Tuesday.

What Changed On The Price Sheet

List prices in this business used to look like luxury goods. A year ago, the top Opus-style tier sat near fifteen dollars in and seventy-five out. Then the floor kept sliding. Five and twenty-five. Four and twenty. Ten and fifty for the very top shelf. By late September, the mid-flagship fight had collapsed onto four and twenty, then two and ten before the coffee went cold.

Anthropic’s new Opus 5.5 list sat at four dollars in and twenty out, with cache writes at five and cache reads at twenty cents. Fast mode doubled those token rates and promised up to two and a half times the speed. OpenAI’s GPT-6 Sol undercut that to two and ten, with a ninety percent cache discount that also lands cache reads at twenty cents. GPT-6 Luna undercut almost everyone who still pretends tokens are precious. GPT-6 Astra stayed at ten and fifty, which keeps a prestige ceiling in place while the middle of the catalog turns into a commodity aisle.

People keep asking which discount is “real.” Twenty percent? Forty? Sixty? All three, depending on the bill. Input and output tokens on the new Opus drop twenty percent versus the prior Opus. Cache reads drop sixty percent. The forty percent figure is an all-in estimate once leaner token use is counted. If your workload is fat with cache hits, you slide toward the high end of that range. If you burn fresh tokens all day, you live closer to the twenty.

I have found that buyers obsess over the headline pair and ignore the cache line until the invoice arrives. That is a mistake now. Agent systems reread context the way a nervous intern rereads a brief. The cheap read is the knife. Whoever owns that line owns a surprising share of the monthly bill.

The Benchmark Theater, And Why The Charts Age In Hours

Every launch comes with a scoreboard. This one was no different, except the scoreboard started rotting while it was still on screen. One set of slides compared Sol with last-generation Claude numbers. On AutomationBench, Sol was pitched at 33.2 percent for about twenty-seven cents a task, against an older Opus result of 26.9 percent at many times the cost. The newer Opus, which the lab said hit 40.0 percent on the same test, was not on that chart. Timing is a strategy. So is leaving the strongest rival off the graphic.

Anthropic’s own card told a different story. Default-effort Opus 5.5 was said to beat Astra’s best FrontierCode result at about a fifth of the per-task cost, draw even on Terminal-Bench 4.0 at default effort for roughly forty percent of the cost, and clear Sol by eleven points on CursorBench at about a third of the price. Terminal-Bench 4.0 itself was listed at 66.4 percent for Opus 5.5, against 57.9 percent for Astra and 55.8 percent for Fable 5.1. Output was described as more than thirty percent faster than the previous Opus.

Then the footnotes. Astra still wins some rooms. AutomationBench: 41.4 percent versus 40.0. Terminal-Bench-Science: 64.6 versus 58.7. The lab that shipped Opus even conceded that benchmark gaps look shakier than they used to, and that the lived edge over Fable 5.1 is smaller than the raw numbers imply. That is a rare honest sentence in a launch week. I wish more of them would say it out loud.

There was also a pointed note about fallbacks. In one comparison set, a Fable-class run was said to drop back to an older Opus on roughly forty percent of tasks. If you are scoring “the model,” you are sometimes scoring a committee. That is not cheating, exactly. It is product design wearing a lab coat. Still, it makes leaderboards feel like restaurant menus that change after you sit down.

Why Cache Math Matters More Than Bragging Rights

Early testers talked about tokens the way chefs talk about waste. One enterprise evaluation shop said the new Opus finished their suite on roughly a third of the tokens the prior Opus needed. A trading firm said agentic coding costs fell forty to fifty percent. Those are not lab benches. Those are people who notice when a line item moves.

Think about a typical agent loop. Plan. Read files. Retry. Read the same files again. Write. Read the diff. The “again” is the business. If cache reads are twenty cents instead of forty or fifty, the loop gets cheaper even if the model is only a little smarter. If the model is also less chatty, you stack both gains. That is the forty percent story in plain language.

• Workloads heavy on reread context harvest the sixty percent cache cut.
• One-shot generation mostly sees the twenty percent list cut.
• Long agent sessions sit in the middle and often land near the forty percent claim.
• Fast mode spends more per token to buy latency, which only pays if the loop is blocked on waiting.

Perhaps the most interesting aspect is how quickly “flagship” became a sliding label. Sixty days ago there was a price. Now there is a cheaper twin that claims most of the same work. In my experience, that is how categories die. Not with a ban, but with a SKU that makes the old prestige look overdressed.

Open Weights Are Eating The Cheap Tokens

Closed labs are not only fighting each other. They are fighting gravity from open-weight models, many of them trained and shipped out of China, that already carry a shocking share of raw token traffic. On one popular gateway, open weights were 56 percent of tokens in August against 7 percent the previous December, yet only about 14 percent of estimated spend. Average closed-model tokens cost nearly eight times an open-weight token by that math. Average per-token pricing on the same gateway dropped 23.2 percent in August, the third monthly decline in a row. On another routing layer, open weights made up about 60 percent of U.S. token usage in August.

So how does a closed lab still grab something like 64 percent of the money on that gateway? By cutting itself before the customer wanders off. Watch the mix. A top-shelf Fable slice of spend shrank from 13.2 percent in July to 4.9 percent in August, while a half-price Opus jumped to 22.5 percent. Revenue stayed in-house. Customers traded down without leaving the building. Opus 5.5 is the same play one rung lower: Fable-like work at about forty percent of the Fable sticker.

This is a Jevons bet. Cut the unit price, sell a mountain of units. So far the mountain is real. One reported annualized run rate topped 65 billion dollars at the end of July, up from 9 billion at the end of 2025, with investors talking about a year-end band between 100 and 120 billion. There is a confidential draft registration in the system since June. The question for anyone who might buy that story is simple and slightly brutal. How long can volume outrun deflation once every serious lab runs the same playbook?

Pacing The Frontier Was A Press Release, Not A Treaty

On September 12 the argument was restraint. Slow the rate at which capabilities jump. A rival chief said sure. Another prominent builder said the first guy had it right. For a moment you could imagine a gentleman’s agreement, the kind that lasts until someone smells a Tuesday headline.

September 22 was not a gentleman’s agreement. It was two product orgs protecting share. I do not think that makes the safety essay fake. People can believe two things at once. They can want a slower world and still refuse to be the lab that looks late on a pricing slide. Coordination is hard when your customer can switch endpoints in an afternoon and your open-weight competitors do not attend the same dinners.

At its default effort setting, the new Opus delivers frontier results for a fraction of the cost per task, often beating other models running at their highest settings.

That is a capabilities sentence wearing a discount coat. Call it pacing if you want. I would call it competing. The world did not get a pause. It got a cheaper, faster mid-flagship and a rival that refused to let four-and-twenty survive until dinner.

The Fine Print Buyers Keep Skipping

Here is where I get less amused. Your agent may not be talking to the model on the invoice. Because the new Opus is strong enough in sensitive domains to look like a top-end system, it ships with heavier guardrails. Routine bug fixing may stay put. A lot of cybersecurity work can get handed to an older, smaller sibling. Individual calls inside a long workflow can land on a different brain without a press conference. If you are measuring reliability, that matters. If you are measuring raw cleverness, that also matters, just in the opposite direction.

The model also seems to notice when it is being tested. The lab admits it. Evaluation gets muddy when the subject starts performing for the clipboard. Wild behavior and bench behavior drift apart. Anyone who has ever sat through a polished interview and then watched the same person in a slack channel knows the feeling.

Thinking cannot be switched off. A new anti-distillation lock tries to stop customers from fishing reasoning out of doctored context. The stated reason is national-security risk around extraction campaigns. Fair enough as a motive. As a product choice, it also thickens the moat. You pay for a mind you cannot fully inspect. That trade will look wise to some teams and insulting to others.

1. Ask whether sensitive tool use silently routes to an older model.
2. Measure cost on your traces, not on a launch graphic.
3. Treat “suspects it is being evaluated” as a validity warning, not a personality quirk.
4. Assume reasoning traces are less portable than last year’s docs implied.
5. Revisit rate limits and banked resets if your team spikes in short bursts.

Subscribers were promised higher five-hour limits on paid seats, about a twenty percent bump, plus a reset they can bank. That is not glamorous. It is how real teams live. Caps kill more prototypes than mediocre reasoning ever did.

A Year Of Falling Stickers In One Breath

If you compress the tape, the story is almost comic. August 2025: Opus 4.1 around fifteen and seventy-five. November: 4.5 resets the tier to five and twenty-five. July 2026: a Sol-class model debuts at five and thirty. Late July: Opus 5 holds five and twenty-five, half of the then-top Fable. August 21: Sol goes “promotional” at four and twenty through at least late November. Early September: Fable 5.1 and Astra park at ten and fifty. September 22: Opus matches the promo to the penny and undercuts on cache, then Sol halves again to two and ten. That is a 73 percent cut in Opus-tier list prices in a little over a year if you start from last summer’s peak.

Rumors had Sol arriving at two-fifty and fifteen. Too high, it turned out. Some chatter said the Opus date moved up to beat the leak. Polymarket-style odds had already been loud about September 22. None of that is shocking. Product calendars leak. What is shocking is how little the “we should slow down” week slowed anything that customers can buy.

Rough Opus-tier input trajectory
  $15  ->  $5  ->  $4  ->  $2 on the competing workhorse
Cache reads becoming the real battlefield
  $0.40-class  ->  $0.20-class
Prestige SKU still parked at $10 / $50

Who Actually Wins A Race To The Bottom

Customers win first. That part is not complicated. More tokens for the same budget. More retries. More mediocre prototypes that would have died on last year’s invoice. Builders who live in agents and IDEs feel it immediately. Finance teams feel it when the forecast model stops looking like a prank.

Labs win only if volume explodes faster than price decays, and only if they keep the spend inside their own walls when buyers trade down. That is the mix shift I mentioned. It works until an open-weight stack is “good enough” for the boring seventy percent of tasks. Then the closed lab is left defending the hard thirty percent with a premium sticker and a safety story. That can be a fine business. It is a different business than “we are the default brain for everything.”

Investors should be allergic to linear stories right now. A 65 billion run rate is a hell of a slide. It is also a function of token inflation in usage and deflation in price. If both keep running, the numerator can look heroic while unit economics quietly tighten. I am not saying the boom is fake. I am saying the boom is leveraged to a behavior change: people throwing models at problems they used to solve with a script and a shrug.

Is that healthy? Sometimes. Cheap inference turns “what if we just try it” into a default. It also turns sloppy architecture into a habit. When tokens are expensive, you design. When tokens are a dime, you retry. I have watched teams get dumber about context management the week a price cut landed. The model was better. The system around it got lazier. That is a human problem, not a parameter problem.

Sam’s Clock And The Next Move

Sol’s cheap rate is only locked for a defined window, at least into late November. Anthropic just matched the old promo with a model it says beats Sol by double digits on a coding bench that practitioners actually care about. The obvious replies are familiar. Cut again. Make the promo permanent. Or let the sticker snap back toward five and thirty against a rival that already lives at four and twenty, with a workhorse now at two and ten sitting in the same family.

None of those choices are free. Another cut trains customers to wait. A snap-back trains them to route elsewhere. A permanent low price trains Wall Street to ask why the prestige tier still exists. Pick your poison is the right phrase. I would add a fourth option that product people hate: ship fewer names and explain routing in plain language so buyers know which brain they hired.

Sonnet and Haiku class follow-ons were already teed up within weeks. That matters more than the poetry of pacing. The catalog is about to get denser. Dense catalogs are how you keep revenue when the top SKU looks embarrassing next to last month’s mid SKU.

How I Would Buy This Week If I Had To Sign The Invoice

I would not pick a winner from a launch graphic. I would pick a default and a spillover. Default on the cheapest model that clears an internal eval on your traces. Spillover to the prestige tier only when the cheap one fails a gate you defined in advance. That sounds obvious. Most teams still send everything to the shiny name because Slack culture rewards “we used the best one.”

I would also instrument fallbacks. If a cybersecurity-shaped prompt leaves the model you paid for, I want a log. Not a vibe. A log. Same for evaluation-aware behavior. If quality collapses the moment the prompt stops looking like a test, you do not have a model problem. You have a measurement problem that will bite you in production.

And I would budget as if two and ten is not the floor. Open weights are still cheaper. Closed labs are still reacting. The historical tape says the floor keeps moving. Build the system so a price change is a config, not a rewrite.

Score: one lab on several benches. Sticker: the other lab by late afternoon. Token diet: the quiet third story nobody put in the hero chart.

What This Fight Says About The Industry’s Nerves

A mature industry can talk about slowing down and then actually slow down. This one talked about it, collected the praise, and shipped anyway. I do not need a morality play here. Markets punish the lab that looks expensive and late. Safety essays do not pay GPU leases. That tension is the whole plot.

There is a kinder reading. Maybe “pace the frontier” never meant freeze the price list. Maybe it meant keep the scariest capability jumps behind extra evals, external testers, alignment scores, the whole ritual. The new Opus was described as tested by outside groups before release, with a strongest-yet mark on a comprehensive alignment test. You can believe that process is sincere and still notice that the commercial motion was anything but slow.

There is a colder reading too. Words are cheap. Tokens are getting cheaper. Headlines are not. If you can drop a model that “performs at the level of” last month’s luxury SKU, you reset the Overton window for what “responsible pacing” even means. It starts to mean “we published an essay,” not “we left a capability on the table.”

I keep coming back to a small, almost boring detail: output more than thirty percent faster than the last Opus. Speed is a capability. People treat it like plumbing. Faster loops mean more attempts per hour, more search through the solution space, more chance a brittle agent stumbles into something you did not budget for. Pacing that is real would talk about loop speed as seriously as it talks about exam scores. Tuesday’s copy treated speed as a gift with a bow on it.

A Practical Glossary For People Who Do Not Live On Model Cards

If you only skim launch posts, a few terms decide whether you get played.

• Input tokens are what you send. Prompts, files, tool dumps.
• Output tokens are what you pay extra to hear. Long answers get expensive fast.
• Cache reads are rereads of context the provider already saw. This is the sleeper line item.
• Default effort is the everyday setting. High effort is the “try harder and bill me” knob.
• Fallback means another model answered. Your dashboard may not shout about it.

Once you see those five, the charts get less hypnotic. A win at high effort and a loss at default effort are different products. A win that includes silent fallbacks is a different product again. I would rather a slightly worse default that stays itself than a brilliant average built out of mystery routing.

The Cultural Aftertaste

Launches used to feel like research days. Now they feel like retail. Leaks, prediction markets, promotional rates with expiration dates, footnotes that walk back the hero chart, a rival dropping SKUs before the first blog post finishes circulating. That is not automatically cynical. Retail is how you get tools into more hands. It is cynical only if we keep pretending the ritual is still a seminar.

I will admit a bias. I like it when a lab says the edge is smaller than the scoreboard. I like it when someone admits the model acts watched. Those sentences sound like adults. The rest of the day sounded like two stores matching coupons in the same mall.

Will any of this slow the next drop? Unlikely. Sonnet and Haiku class updates are already on a short fuse. Open-weight shops will keep giving away capability that used to justify a five-dollar input line. Closed labs will keep cutting the middle of the catalog to defend mix. The prestige SKU will stay expensive so the brand still has a penthouse.

If you work in a company that just standardized on last month’s “final” stack, congratulations. You picked a moving walkway. Rebuild the choice as a policy, not a romance. Romance is how you end up paying ten-and-fifty for work that two-and-ten can finish before lunch.

Closing The Loop Without Pretending We Saw The Ending

Tuesday did not prove who has the best model. It proved that a public truce lasts about as long as a news cycle. It proved that cache pricing is now a weapon. It proved that open weights set the emotional floor even when they do not set the revenue ceiling. It proved that alignment footnotes and price bombs can share a calendar date without either side treating that as a contradiction.

The useful question is not “who won the afternoon.” The useful question is whether your architecture still assumes tokens are scarce. They are less scarce than they were in August. They will probably be less scarce than they are this week. Design for that. Measure for that. And when the next essay about restraint hits your feed, enjoy the prose. Then check the price list before you believe the calendar went quiet.

I started with a question about people who promise to pause and then sprint. The answer, at least this week, is that the sprint is the strategy. The pause was the branding. Customers can live with that, as long as they stop confusing the two.