Let’s appearance at a quick timeline of the important LLM example announcements from the two OpenAI and Anthropic complete the former few months. All pricing numbers provided are input MTok / output MTok / cache peruse MTok.
June 26, 2026#
OpenAI announced GPT-5.6 Sol and Luna. Sol is the straightforward successor to GPT 5.5 at the identical cost of $5 / $30 / $0.50, during Luna is a small example specifically designed for cheaper use cases at $1 / $6 / $0.10. They additionally announced a Terra example that’s in between those two models, but few cared concerning it.
June 30, 2026#
Anthropic announced Claude Sonnet 5, which Anthropic claimed to be nearer to their Opus models, but feedback was mixed. Notably, it had initiate pricing of $2 / $10 / $0.20 which was initially expected to be impermanent but would go up.
July 24, 2026#
Anthropic announced Claude Opus 5, which was a long-awaited major-version update stated the impact Opus 4.5 had on the increase of agentic coding, at a cost of $5 / $25 / $0.50 (similar to GPT-5.6 Sol). Although several of the provided benchmark results amazingly exceeded those of the much larger Fable 5, real-world feedback was additional mixed.
July 30, 2026#
OpenAI announces that they cut the cost of GPT-5.6 Luna by 80% (???) to $0.20 / $1.20 / $0.02.
Model norm vs. Cost-per-task for GPT-5.6 Luna following the enormous cost reduction. Top-and-left is better, and if OpenAI didn’t do the 80% cost cut, it would have misplaced to DeepSeek V4 Pro. via OpenAI
August 10, 2026#
Anthropic announces the impermanent Sonnet 5 cost cut for the initiate is now permanent.
August 21, 2026#
OpenAI announces that they are cutting the cost for GPT-5.6 Sol by ~20% ($4 / $20 / $0.40) for the next three months.
September 1, 2026#
Anthropic announced Fable 5.1 at a cost of $10 / $50 / $0.25, which, in supplement to having strictly improved benchmark achievement in all areas than Fable 5 from June at $10 / $50 / $1, cut the cache peruse disbursal by 75%.
When using agents for coding, cache peruse expenses are the huge bulk (96% - 99%) of the total token expenditure as the cache is what contains the complete former conversation all turn. Additionally, the cache tends to be hit about 90 - 100% of the period (varies by the infrastructure of the LLM conclusion provider). Taken together, a 75% decrease in the cache peruse disbursal correlates to a 70-75% decrease in the disbursal of using the model—obviously notable, equal if cache expenses are typically a footnote!
September 3, 2026#
OpenAI announced GPT-6 Astra, which is akin in size to Fable 5.1 and has akin pricing at $10 / $50 / $1 but without the lesser cache peruse cost. Astra’s example achievement is certainly comparable to Fable.
September 22, 2026#
OpenAI announced GPT-6 Sol at at a shocking cost of $2 / $10 / $0.20 that halves the input and output cost related to the temporarily-discounted GPT-5.6 Sol! They simultaneously released GPT-6 Luna additionally at fractional the input and output cost ($0.10 / $0.50 / $0.01) of the former version.
On the same exact day, Anthropic announced Claude Opus 5.5, anywhere it’s explicitly stated “Opus 5.5 requires small compute to assist than Opus 5, and its pricing reflects that”. Its cost of $4 / $20 / $0.20 is much lesser in total than Opus 5’s because of a 20% input/output disbursal cut and a larger cache peruse cut, resulting in a total disbursal decrease of 40-60% for a example that’s equal improved than Fable 5.1 (I have fairly a lot of things to say concerning Opus 5.5, but that’s for another post).
September 29, 2026#
It turned out that GPT-6 Sol was a dud performancewise, so a week afterward OpenAI announced GPT-6.1 Sol, which amazingly has comparable achievement to Astra (can verify following individual use), but OpenAI casually cut the cache cost in fractional ($2 / $10 / $0.10) matching Anthropic’s 5%-of-input cache peruse cost, making this example family cumulatively 75% cheaper.
Overall, starting from the publish of the GPT-5.6 example families in June to now, the disbursal of using these famous LLMs has cratered, which is not item I would have expected a twelvemonth ago. These cost cuts additionally use to example use from subscription agentic harnesses specified as Codex and Claude Code. More importantly, these models do not appear to have cut any corners in conditions of output norm to accomplish their cost cuts.
Data from Vercel’s AI example gateway supporting the decreasing cost-per-token trend to 43.3% of the April 2026 cost. However, total expend has been expanding since the models are getting improved faster than they’re getting cheaper! via Vercel
From the timeline above, there is a apparent tit-for-tat between Anthropic and OpenAI, and it is sensible to assume that these particular cost cuts from OpenAI and Anthropic are in reply to all other. However, no one cuts prices by 80% unless there is several externality in play. So, what gives?
The Behavioral Economics of LLMs and Competitor Pricing Pressure#
If you have been evaluation agents alongside new LLMs as they are released, it can be extremely difficult to differentiate them exterior of subjective vibes. Since final November that launched the hypergrowth of agentic LLMs, models have rotate into additional of a commodity anywhere it’s trivial to try a new example on your codebase alongside fair a toggle in a harness akin OpenCode and Zed Agent, or a distinct API endpoint cord through OpenRouter, and hence there is small moat. Additionally, there are many competitors—not fair ample companies akin Google and Meta—who can provision models which are sufficient substitutes and hence authorize pricing pressure.
For open importance models, let’s obtain another appearance at DeepSeek Flash V4.1, which additionally lately released an updated 1017 variant.
OpenRouter at this item is the location you go for open-weight models. Retrieved October 24th, 2026.
The first DeepSeek V4 Flash 0420 was released in April 2026 at a cost of $0.14 / $0.28 / $0.0028 and went megaviral as a result, particularly as OpenAI didn’t have a small-class example for uncomplicated tasks at the time. That $0.0028 turned out to be the key, as a cache peruse proportion of 2% is substantially lesser than the 5%-10% of the OpenAI and Anthropic models above, which makes the productive disbursal even lesser than what is advertised. I doubtful the 80% GPT-5.6 Luna cost cut in July was meant to specifically vie alongside these smaller open-weight models. Now, the new DeepSeek Flash V4.1 1017 example performs significantly improved than DeepSeek Flash V4 benchmark-wise (45 on the Artificial Analysis indicator vs. 24) and the disbursal from DeepSeek remains akin at $0.15 / $0.60 / $0.003. Note the operative words “from DeepSeek”.
On OpenRouter, any conclusion provider can catalog a example that they host, and whenever a person hits the OpenRouter API, the petition is routed to one of those models—it’s routed to any accessible example by default for burden balancing but the developer can specify preferences specified as “lowest price” in the API request. Back in May, I noted that there were 13 distinct providers hosting DeepSeek V4 Flash on OpenRouter. For the updated 1017 variant, there are now 29 distinct providers.
Providers for DeepSeek V4.1 Flash 1017 sorted by cache peruse cost. Retrieved October 24th, 2026.
However, looking at the productive input expenses (which document for the communication between cache peruse expenses and cache hit rates), many distinct providers are near costwise.
Effective expenses by endpoint, alongside Cloudflare and DeepInfra inside 10% of DeepSeek’s cost. Despite the fact that these models are the cheapest, note that these models lone have a smaller group of the total token share. Retrieved October 24th, 2026.
OpenAI and Anthropic volition have to differentiate from these open importance models alongside higher-quality models—they are currently doing a fine job alongside that particularly alongside Opus 5.5—but they need to keep doing that, which is why billions of dollars are being invested into R&D. Open-weight models are catching up fast: fair a few days ago, Moonshot AI announced Kimi K3.1 Pro which has comparable achievement to Opus 5.5 (Kimi at 55 vs. Opus 5.5 at 58) at fractional the input/output cost and the identical cache peruse disbursal ($2 / $10 / $0.20). It’s the identical size and architecture as Kimi K3 which is currently being hosted by complete 20 distinct conclusion providers: formerly the weights autumn on November 4th, you can hence anticipate most of them to assistance the new Kimi K3.1 Pro example at a approximately equal price. Notably, Moonshot AI itself is joining the cost cut fun and priced Kimi K3.1 Pro at small than the first Kimi K3 ($3 / $15 / $0.30).
One recurring trend among vocal AI detractors is that the cost of tokens has been kept artificially low to hook group into the LLM ecosystem, but afterward the prices volition addition to “the true disbursal of a token” and the AI boosters volition be screwed formerly the liberated journey ends. The theory is that OpenAI and Anthropic are loss-leading fair akin what Uber and Lyft did to the ridesharing industry in the 2010s, but dissimilar then, OpenAI and Anthropic are not a duopoly who have the sole power to set prices in the ecosystem and the obstacle to admission for the conclusion industry is not as high. As lengthy as these many conclusion providers can continue to assist models at a akin cost and a akin norm to what OpenAI and Anthropic can provide, and switching disbursal is negligible, afterward the prices may continue to decrease more. It’s no astonishment the two companies are trying to build moats to defend against this trend, specified as a concentration on enterprise commitments and individual agentic assistants akin OpenAI Dots and Claude Duet.
The idea that token prices volition abruptly increase is additional a imagination that is not supported by the two behavioral economics and the genuine trends of LLM prices going downward significantly complete the final few months for improved models, and not fair from OpenAI and Anthropic. As lengthy as scalers can cost at a cost-per-token lesser than the revenue-per-token, afterward those conclusion providers volition decrease token expenses to vie alongside all another to a item that OpenAI/Anthropic volition have to lesser their prices to vie too. The 29 distinct providers on OpenRouter have small incentive to assist an unprofitable example (they would fair halt serving it and use the compute capability for a profitable example instead) and “the true disbursal of a token” may end up being anywhere cost equals marginal cost.
On that note, the two OpenAI and Anthropic are certainly near to IPOing, which raises additional questions concerning the cost cuts.
IPO Pumping#
The timing of these cost cuts is notable since the two OpenAI and Anthropic are rumored to be IPOing soon—OpenAI in 2027, and Anthropic inside a brace months. So far, leaked financials from the two OpenAI and Anthropic show unsurprisingly that they are losing billions because of R&D/scaling, albeit in the two cases these reports do not document for the hyperexpansion of agentic LLM use complete the former few months. Investors volition scrutinize the enormous negative profits and petition that they have a scheme to rotate things rear around.
So, why did they cut their chief income origin by up to 80%? That’s the opposite of what you would desire to do to appearance fine on an S-1 prospectus! The gain margins on conclusion themselves apt inform a distinct story, alongside several reports saying the profit-per-token margins are as fit as 80% alongside caveats. How much of a gain border do they have following these important cuts?
Speaking personally, the cost cuts have certainly affected how I expend on LLM agents, but not in the way I anticipate OpenAI/Anthropic desire their representative client to behave. I used to have a $100/month subscription to Codex (the 5x Pro plan) whenever GPT-5.6 Sol was the top model, and I managed to fair completely exhaust all the use included in the subscription notwithstanding getting annoyed alongside random use resets. Once GPT-6.1 Sol released and I noticed that the disbursal of using it was efficiently cut by 75%, you cognize what I did?
I cut the max use quota to 1/5th following getting 4x additional productive use of the identical model, so it ends up approximately neutral and the apparent choice.
In fairness, the gain margins for subscriptions specified as Claude Code and Codex are uncertain stated their huge value related to the API usage, and if there are chief crunches I doubtful subscription value volition be cut before the basis cost of the models. For example, at the identical period GPT-6.1 Sol was announced, OpenAI cut the use allowance by 50% for the $200/month tier in Codex under the disagreement “to indicate our increasingly productive models”—which is technically accurate but it’s comic to put in writing.
I volition be very inquisitive to see the Anthropic IPO prospectus and the conclusion gain margins formerly it becomes community soon.
I additionally inquiry whether it’s fine for OpenAI/Anthropic if they intentionally obsolete their additional costly models. After Opus 5.5 released a duration ago, there hasn’t been any conversation concerning either Fable 5.1 or GPT-6 Astra since Opus 5.5 is fair that good.
Top-and-left. via Anthropic
I completely anticipate a Fable 5.2/GPT-6.1 Astra to publish in a brace weeks and demonstrate me wrong, and I volition be very inquisitive to see how it volition be priced.
Another less-cynical aspect that could explain the cost cuts is that LLM infrastructure has improved complete the former few months—when Anthropic stated “Opus 5.5 requires small compute to assist than Opus 5, and its pricing reflects that,” I accept it. One outcome that would be genuinely ironic is that if these models are additional compute efficient, it reduces the need for these enormous data center buildouts, which could logic cascading financial failures. However, that apt won’t happen since conclusion is lone one part of the compute mix and those data centers are motionless needed for another parts of the pipeline.
Overall, the cost cut is weird, but as a person I’m certainly not complaining! Open-weight models and secondary conclusion providers volition continue to assist as an financial inspect on the prices of LLMs and another endeavor shenanigans the ample labs may concoct. Even if Anthropic’s and OpenAI’s IPOs are catastrophic and logic the two companies to fold, it volition not abruptly logic LLMs and agents as an industry to die.
