Connect with us

NEWS

LLM Prices Crash 80% While Total Spend Keeps Climbing

OpenAI and Anthropic cut token prices by up to 80%, yet Vercel data shows total spend rising as Codex allowances shrink on October 30.

Published

on

OpenAI and Anthropic have cut list prices on their main models by as much as 80% since June. The cheap-token era did not shrink the checks. Vercel’s gateway shows September spend about 13 times December’s, while the average token cost 47% less, and on October 30 ChatGPT Pro’s $200 plan loses half its Codex allowance.

Cache reads, not the headline input rate, now set the bill for coding agents. Open-weight hosts keep pushing that rate down, and the labs are answering with cheaper mid-tier models plus tighter subscriptions.

Four Months of Cuts, Down as Much as 80%

The first move was a small model. On July 30 OpenAI cut GPT-5.6 Luna from $1 / $6 / $0.10 per million tokens to $0.20 / $1.20 / $0.02, an 80% drop on the list. Anthropic then locked in Sonnet 5’s launch rate and, on September 22, shipped Claude Opus 5.5 at $4 / $20 / $0.20. That same day OpenAI halved GPT-6 Sol and Luna against their GPT-5.6 promo rates. A week later GPT-6.1 Sol arrived at $2 / $10 / $0.10 cached.

THE CUTS SINCE JUNE

  1. June 26, 2026: OpenAI ships GPT-5.6 Sol at $5 / $30 / $0.50 and Luna at $1 / $6 / $0.10.
  2. June 30, 2026: Anthropic launches Claude Sonnet 5 at $2 / $10 / $0.20.
  3. July 24, 2026: Claude Opus 5 lists at $5 / $25 / $0.50.
  4. July 30, 2026: Luna’s 80% cut lands.
  5. August 21, 2026: GPT-5.6 Sol promo pricing falls to $4 / $20 / $0.40.
  6. September 1, 2026: Fable 5.1 keeps $10 / $50 and cuts cache reads 75% to $0.25.
  7. September 22, 2026: GPT-6 Sol and Luna take another 50% off, and Opus 5.5 ships at $4 / $20 / $0.20.
  8. September 29, 2026: GPT-6.1 Sol matches Sol’s $2 / $10 and halves cached input to $0.10.

OpenAI now sells that mid-tier as near-Astra intelligence for a fifth of the price. Astra still lists at $10 / $50 / $1. Cached input on GPT-6.1 Sol is 95% below its $2 uncached rate, which is the number that matters once an agent starts rereading its own context. On a hard science-agent test, OpenAI put GPT-6.1 Sol at $5.47 per task against $23.21 for Opus 5.5 and $23.80 for Astra. Factual errors on a stacked ChatGPT set fell from 11.4% to 7.7%.

CURRENT LIST PRICES PER MILLION TOKENS

Model Input Output Cache read
GPT-6 Astra $10.00 $50.00 $1.00
Claude Fable 5.1 $10.00 $50.00 $0.25
Claude Opus 5.5 $4.00 $20.00 $0.20
GPT-6.1 Sol $2.00 $10.00 $0.10
Kimi K3.1 Pro $2.00 $10.00 $0.20
GPT-6 Luna $0.10 $0.50 $0.01
DeepSeek V4.1 Flash (off-peak) $0.15 $0.60 $0.003

GPT-5.6 Sol’s promo rates remain posted at least through November 21. Nobody is using them as the working mid-tier anymore.

Cache Hits Carry Almost All of an Agent Bill

A coding agent resends the system prompt, the tool list, and the repo map on every turn. Cache reads make up 96% to 99% of that token pile, and hit rates sit around 90% to 100% depending on the host. Cut the cache rate 75%, as Anthropic did from Fable 5 to Fable 5.1, and the session cost falls by about 70% to 75% even if input and output barely move.

That is why GPT-6.1 Sol’s real change was the cached line, not the $2 / $10 pair it already shared with GPT-6 Sol. Anthropic did the same trick on Opus 5.5: input and output fell 20% from Opus 5, cache reads fell 60% from $0.50 to $0.20, and the lab said typical workloads land about 40% cheaper. Fast mode doubles those rates to $8 / $40 for up to 2.5 times the speed.

Heavy users already treat Opus as an advisor and dump file search onto cheaper siblings. The list price on the flagship is not the unit they pay. The unit is cache on the model they actually loop.

Why the Bill Went Up After the Price Fell

Vercel’s AI Gateway, which routes tens of trillions of tokens a month for production apps, is the cleanest public meter of this. Its October index, covering September traffic, found open-weight models took 71.5% of tokens and 23% of spend. September spend ran about 13 times December’s. The average price per token was 47% lower.

VERCEL GATEWAY, SEPTEMBER TRAFFIC

  • Open-weight volume: 71.5% of tokens, up from 7% in December 2025.
  • Open-weight dollars: 23% of spend, up from about one in twenty in June.
  • Total spend: about 13 times December’s level.
  • Unit price: 47% below December, after a 23.2% drop in August alone.

Closed models still take most of the money. In August, the month before those September figures, Anthropic held 64% of gateway spend and had not fallen below 61% in any month since December. Fable 5’s share collapsed from 13.2% in July to 4.9% in August after Opus 5 arrived at roughly half the token price and rose to 22.5%. Nine in ten Fable teams cut usage. They did not leave Anthropic. They stepped down a tier.

Astra, launched September 3 at Fable 5.1’s $10 / $50, took a third of OpenAI’s gateway spend within two days and 7.7% of all gateway spend in its first twelve days, against 3.7% for Fable 5.1. Luna processed more than eight times as many OpenAI tokens as Astra in that window and generated about one-ninth the spend. Cheap models carry the volume. The dollars still sit on the models teams will not risk being wrong on.

Open-Weight Hosts Race Under DeepSeek’s List Price

DeepSeek V4 Flash went megaviral in April at $0.14 / $0.28 / $0.0028, when OpenAI still lacked a small everyday model. The July Luna cut was the reply. The 1017 snapshot of V4.1 Flash, posted October 17, scores 45 on the Artificial Analysis index, against 24 for the original V4 Flash, and DeepSeek still charges $0.15 / $0.60 / $0.003 off-peak and $0.30 / $1.20 / $0.006 at peak.

The floor is not DeepSeek’s own sticker. It is the crowd of hosts who serve the same weights. In late October the DeepSeek V4.1 Flash 1017 provider list ran to 29 distinct hosts on OpenRouter. Several sit within about 10% of DeepSeek’s own effective cost once cache hits are counted. Those hosts have little reason to sell an unprofitable model; they would just move the GPUs. So long as they can price above marginal cost and still undercut the labs, the labs have to follow.

Switching cost is a toggle in a coding harness or a different endpoint string. That is the pressure the 80% cuts were answering, and it has not eased.

Moonshot’s Open Weights Arrive on November 4

Moonshot AI put Kimi K3.1 Pro on its API on October 23 at $2 / $10 / $0.20, a third below Kimi K3 and half of Opus 5.5’s list. Artificial Analysis scores it at 55, three points behind Opus 5.5 at 58 and ahead of Fable 5.1, GPT-6 Astra, and Gemini 4 Argon at 53. On Terminal-Bench 4.0, Moonshot reports 65.7% against 66.4% for Opus 5.5 and 58.2% for Astra. The model is 2.8 trillion parameters total, 104 billion active, with a 1 million token context.

The API price is a placeholder. Moonshot says the full model weights will be released by Nov 4. Kimi K3 already sits on more than 20 inference hosts. Once the new files land, those hosts can price K3.1 Pro near Moonshot’s own rate, the way they already price DeepSeek Flash. A 2.8 trillion parameter checkpoint going up like a public file is the constraint on every closed mid-tier from that week on.

K3.1 Pro also writes more tokens per answer than K3, so the cheaper sticker does not fall one-for-one onto cost per task. That is a real drag. It does not change the direction of the floor.

The $200 Codex Plan Loses Half Its Allowance on October 30

API cuts and subscription cuts are not the same gift. OpenAI reopened the $200 ChatGPT Pro plan to new buyers on September 29 with a smaller Codex allotment, and existing subscribers keep the old one only through October 29. On October 30 they move to the new math at the same $200. The multiplier falls from 20 times the Plus allowance to 10 times.

In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.

Thibault Sottiaux, Codex lead at OpenAI, on X

Sottiaux also said subscribers will still get more work done than on the $200 plan a month earlier, because the models got cheaper and better in the same window. OpenAI’s help copy puts it as a cut “to reflect our increasingly efficient models.” Half the included API dollars at half the token price can net out even on a Sol-class loop. It does not net out if the user stays on Astra, which never moved from $10 / $50 / $1.

Power users already responded the way a rational buyer would. When GPT-6.1 Sol made the same work about 75% cheaper on cache-heavy runs, some cut their own usage cap instead of running five times as many agents. Spend stayed flat. The lab’s volume story still works on the API side, where a lower unit price meets more tasks. The subscription side is where OpenAI clawed back the surplus the $200 plan had been leaking.

Anthropic Raised Caps as OpenAI Tightened Them

Anthropic went the other way on the same week Opus 5.5 shipped. It said the new model requires less compute to serve than Opus 5, cut the list, and raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. It also added a bankable rate-limit reset. In an internal HAProxy rewrite from C to Rust, Opus 5.5 finished in 9.5 hours against 12 for Fable 5.1 and cost 51% less, with both passing nearly all of HAProxy’s own tests.

Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.

Anthropic, Claude Opus 5.5 announcement

That sentence is doing two jobs. It explains the cut without calling it a sale, and it tells IPO readers that serving cost, not panic, set the rate card. Both labs are still expected to file: Anthropic on a November rumor, OpenAI in 2027. Leaked drafts have them losing billions on training even as agent usage exploded. A prospectus that shows token prices falling 80% looks ugly until the same pages show spend climbing because the models got good enough to run all day.

The check on any later hike is still the hosts. If 29 shops can serve a Flash-class model near DeepSeek’s cost, and if Kimi’s weights actually appear on November 4, neither lab can quietly walk prices back to “the true cost of a token.” They can keep the dollars by owning the hard calls, as Anthropic did when Fable customers stepped to Opus, and by metering the $200 plan like a prepaid API, which is what October 30 does. The Codex clock on the old 20x allowance runs out that day.

Disclaimer: This article is news reporting and analysis of public model prices, gateway traffic, and company statements. It is informational only and is not investment advice, a recommendation to buy or sell any security, or a forecast of OpenAI’s or Anthropic’s listing terms. Readers who plan trades or allocations around these labs, their cloud suppliers, or related funds should consult a licensed financial adviser. Prices, usage allowances, and IPO timelines are those published by the companies and data providers named above and can change without notice.

Harry is the editor and publisher of MY WORLD NEWS 24, an independent title under his own ownership. Ten years of reporting and then editing taught him that a global readership is not served by assuming everyone lives in the same country. Stories here state currencies, units and time zones explicitly, name the country a law or a company belongs to, and explain local context rather than treating it as known. That care extends to sourcing: a claim is anchored to the filing, statement, transcript or dataset that made it, wherever in the world it was issued, and each figure is checked against that source before publication. The site reports news, business and technology, science and sports, entertainment, lifestyle and travel, and auto and gaming, all with the same standard of evidence. Mistakes are fixed under a corrections policy anyone can read, and the page carries a note saying what was changed. Harry reads every message sent by readers and replies from support@myworldnews24.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending