Since 17 August, DeepSeek has charged twice as much for the same model during peak hours as it does off-peak. Nothing about the model changes. You are simply paying more when everyone else wants it too. That sounds unusual for AI, BUT it is completely normal for electricity. Power companies have priced peak and off-peak for decades, because demand arrives at predictable times, the grid has to be built for the busiest hour, and cheaper nights move the flexible load out of the way.

AI providers have arrived at the same problem. GPUs are expensive, capacity is finite, and demand is not evenly spread through the day. Nor is DeepSeek alone. Z.ai's new Coding Plan meters GLM-5.3 usage in points that stretch further off-peak, and Anthropic sells Opus 5 with a Fast mode at roughly double the price for answers at two and a half times the speed. One charges you less for waiting; the other charges you more for hurrying.

For a person chatting with a model, none of this matters. You ask when you need the answer. For software it matters a great deal. A coding agent reviewing a repository overnight does not care whether it starts at 1am or 3am, and neither does a batch job generating synthetic data or grinding through evaluations. That gives providers a second lever: rather than building capacity for every peak, they can discount the quiet hours and let customers decide what can wait.

When DeepSeek's change landed, it lifted the index 6% in a week and raised the floor for the first time in the index's history, from $0.182 to $0.352 blended. The spread between the dearest and cheapest member halved, from 121 times to 62.

The bigger implication is that the idea of a single list price is starting to break down. The cheapest token may increasingly be the one that can wait.

Subscribe now

The Token Price Index fell to $2.21 this week after Google launched Gemini 3.7 Flash at $0.75/$3.75, replacing 3.6 Flash in the basket 23 days after it joined. It will not stay there.

DeepSeek is raising its API by Monday, and for the first time since this index started, the cheapest model in the basket will cost more. The change is not a straight increase - it’s more nuanced.

What’s the increase look like?

DeepSeek is replacing its flat rate with peak and off-peak billing, with peak covering seven hours of the day:

Even considered on an off-peak basis:

• V4 Flashgoes from $0.182 blended to $0.352, a 93% increase

• V4-Pro goes from $0.57 to $1.06, up 87%

• The gap between the most and least expensive model in the basket, steady near 121x for months, reduces to about 63x

Change but no change.

What the increase does not do is dislodge DeepSeek from the bottom. Flash at $0.352 still undercuts GPT-5.6 Luna at $0.50 and GPT-5.4 Nano at $0.515.

The floor rises; but DeepSeek still wears the “cheapest” crown. The mechanism is the more interesting part. Google is selling Gemini 3.7 Flash at an introductory rate that expires on 31 December. Anthropic tried the same with Sonnet 5, then changed its mind on 10 August and made the introductory price permanent. DeepSeek, which first experimented with off-peak discounts in early 2025, is now bringing time-of-day pricing back for V4 as explicit peak and off-peak tariffs.

A commercial evolution.

API pricing is acquiring another feature of electricity markets: time-of-use tariffs. What you pay depends not only on which model you call and what kind of token you consume, but when you consume it.

That is harder to track and harder to budget. It is also what pricing starts to look like when capacity, rather than capability, becomes the binding constraint.

On 30 July, three weeks after launching the GPT-5.6 family, OpenAI cut the price of two of its three models.

Luna, the cheapest, fell 80%, from $1 and $6 per million tokens to $0.20 and $1.20. Terra, the mid-tier, fell 20%, to $2 and $12. Sol, the flagship, was left alone.

OpenAI called it efficiency gains and pointed to its own Sol model rewriting production code to help fund the drop, but the timing tells the plainer story: a CNBC investigation this month put Chinese models at 46% of US enterprise token usage, and Luna now sits at $0.50 blended, the second-cheapest model in the index, a US model parked down at the floor the Chinese providers usually own. That one set of cuts is the entire move in the index this week. Nothing else in the basket changed price.

The crown and the ceiling

The week’s other headline barely touched the number. Anthropic released Claude Opus 5 on 24 July at $5 and $25 per million tokens, exactly what Opus 4.8 charged, so swapping one for the other in the index changed nothing at all. What changed is the leaderboard.

For the first time since we started the index, the most capable model in the basket is no longer the most expensive one. The capability crown and the price ceiling, which had sat on the same model for months, have come apart.

Put the two together and you get a fair picture of the frontier in mid-2026. In a single week, competition showed up from both directions: OpenAI cutting prices under pressure, and Anthropic shipping a better model at a price it had already set. If you are buying today, both of those are working in your favour.

Subscribe now

On brand, Kimi K has had a makeover.

Artificial Analysis scores K3 at 57 on its Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol, so Moonshot has built one of the three highest-scoring models on an independent benchmark, running 2.8 trillion parameters across a million tokens of context.

AA puts it at about $0.95 per task, below GPT-5.6 Sol and roughly half of Opus 4.8, which means K3 is not expensive next to the frontier. It is expensive next to the Kimi you were using last month.

One wrinkle (pun intended) worth pricing in: K3 is talkative, generating 130 million output tokens across AA’s index against a 63 million median, and output is the dear half of the bill. It costs less per task than Opus 4.8 and still finds more to say than almost anything else on the board.

As the headlines say - is this the end of cheap Chinese AI?

This is the line everyone reached for, which overstates things, since DeepSeek still sets the index floor at $0.182 and has not moved.

The narrower reading is that Moonshot stopped discounting the moment it had something worth charging for.

On 1 September, when Claude Sonnet 5’s introductory rate expires, Moonshot’s best model will cost exactly what Anthropic charges for its middle one.

If you are choosing today - that’s competition working as it should.

Sol-id entry

OpenAI made GPT-5.6 generally available on 9 July after two weeks of being limited to roughly twenty government-approved partners.

The summary:

• Sol, the flagship, landed at $5 per million input tokens and $30 per million output — precisely what GPT-5.5 has been charging.

• Terra at $2.50 and $15, matching GPT-5.4.

• Luna at $1 and $6, new to the basket rather than replacing anything, as the cheap end of the GPT-5.6 range.

Sol is the better model, scoring 59 on Artificial Analysis’s Intelligence Index against GPT-5.5’s 55, and it costs no more to run.

Grok, paper, scissors

What did move the index was Grok 4.5, which xAI shipped a day earlier, on 8 July, at $2 and $6. Set against Claude Opus 4.8, or against Sol, those are low numbers, and xAI was happy to point that out.

Set against Grok 4.3 (the model it actually replaces in the basket) the price offered is a rise from $1.25 and $2.50. On the index’s blended basis, xAI goes from $1.625 to $3.20, and that single substitution accounts for nearly all of this week’s ~3% increase.

The basket now holds 21 members across 10 providers. GPT-5.5, GPT-5.4 and Grok 4.3 drop out this week - superseded, not retired.

Three threads, still running

The pattern of the last few months holds: (1) capability keeps climbing with little price movement at the top end of town, (2) the cheaper tiers are getting less cheap but remain competitive, and (3) the Chinese-led super-cheap options still undercut everyone.

The ceiling returns

The US lifted its export controls on 30 June, and Fable returned to global availability on 1 July at the same $10/$50 it launched with. It rejoins the index as the ceiling, and at $22 blended it pulls the whole geometric mean up with it.

This is a composition change, not inflation. When the most expensive frontier model vanishes and then reappears, the headline swings, and that swing is the index doing its job. It measures the frontier you can actually buy, and for nineteen days you could not buy the top of it.

Also worth noting is the floor-to-ceiling multiplier (which measures the gap between the cheapest and most expensive index members) is now 120.9x, up from 68.7x in June.

The quieter move runs the other way

Underneath the jump, prices at the default tier fell. Claude Sonnet 5 replaced Sonnet 4.6 as Anthropic’s new default. It scores higher on the Artificial Analysis Intelligence Index (53 against 47) and entered cheaper, at an introductory $2/$10 against Sonnet 4.6’s $3/$15.

The launch that hasn’t reached the index

OpenAI shipped GPT-5.6 on 26 June, but only to about twenty government-approved partners. It comes in three variants, each pitched at a different tier of work:

• Sol: the flagship, built for the hardest problems - complex coding, long-horizon planning, and security research. OpenAI’s most capable model to date.

• Terra: the mid-tier workhorse, matching the previous generation’s performance at roughly half the cost. Aimed at everyday enterprise and developer tasks.

• Luna: the fastest and cheapest, for high-volume routine work like summaries and drafting.

None of them are broadly buyable, so none can enter the basket yet.

Availability is the story

That is the pattern worth watching. Fable was pulled and then restored inside three weeks. GPT-5.6 is out but gated to about twenty partners. While the frontier’s most capable models keep moving in and out of reach, the index tracks availability as much as price.

On 25 June, Sam Altman reportedly told staff that OpenAI will stagger the release of GPT-5.6 at the request of the US government. Rather than a broad launch, the model will first go to a limited group of partners, with federal officials approving access customer by customer during the preview.

The request came from the Office of the National Cyber Director and the Office of Science and Technology Policy, as Washington builds out its new pre-release review process for frontier models.

That phrase matters, because this is the second time this month the government has stepped into a frontier launch. Two weeks ago, an export-control directive aimed at foreign-national access forced Anthropic to disable Fable 5 and Mythos 5 for all customers.

The OpenAI version is lighter, a staggered rollout rather than a global shutdown, but the pattern is the same.

Waiting for Fable 5

Fable 5 remains unavailable, two weeks into the suspension, with no clear restoration date. The story has hardened, not softened.

AP reported that Anthropic’s Mythos model identified vulnerabilities in sensitive US government systems within hours during a sanctioned testing exercise with intelligence agencies.

Senator Mark Warner said NSA and Cyber Command chief Gen. Joshua Rudd had described the tool as breaking into “almost all” classified systems, though officials later clarified that identifying vulnerabilities quickly does not mean the model exploited them within that timeframe.

For the index, the takeaways are blunt: (1) The most capable model on the leaderboard, Fable 5, is offline (2) The most anticipated launch, GPT-5.6, is gated.

For now, the frontier we can measure is increasingly the frontier Washington allows.

On Artificial Analysis’s Intelligence Index (v4.1), GLM 5.2 scores 51, +11 points above GLM 5.1 and the highest of any open-weights model. None of that came from a bigger model. The architecture is unchanged from 5.1 (744 billion parameters, 40 billion active); the gains are from training. The context window went from 200K to a full million tokens.

The ranking is where it gets interesting. GLM 5.2 sits fourth overall on the Artificial Analysis leaderboard, behind Claude Fable 5 (60), Claude Opus 4.8 (56) and GPT-5.5 (55). But Fable 5 has been switched off (more below), so among models you can actually call through an API today, GLM 5.2 is third.

The suitability question is separate from the price one. GLM 5.2 is a Chinese open-weights model, and data residency and procurement rules rule it out for plenty of buyers. The index tracks what frontier capability lists for, not whether you should run it.

This is the same pattern as Opus 4.8 last month: better than 4.7 across coding, agentic and reasoning benchmarks, launched at the identical $5/$25 Opus has held since 4.5.

Watching brief: Fable 5

Fable 5 is still dark. The US export-control directive that pulled it and Mythos 5 offline worldwide on 12 June remains in force, with no restoration date.

This week the negotiation surfaced.

Anthropic sent senior technical staff to Washington and met Commerce Department officials on 15 June, while technical discussions continued through the week. It is unresolved.

More than 80 cybersecurity executives signed a 14 June open letter to Commerce Secretary Lutnick and National Cyber Director Cairncross, arguing the controls stripped frontier tools from cyber defenders.

At the G7, leaders and diplomats discussed a “trusted partners” access framework with Lutnick; a US official told the New York Post that exempting allies would be “completely illogical.”

For the index, nothing changes yet. Fable stays held out as a suspended member of record.

Until then the index measures what is buyable, and the practical ceiling is Opus 4.8 and GPT-5.5.

The week’s irony: the most capable model on the leaderboard is the one nobody outside Anthropic can touch, which is exactly why an open-weights model is now the third-best thing you can buy.

A short update on this week’s issue. Hours after Claude Fable 5 entered the index, the US government issued an export-control directive suspending all access to Fable 5 and its sibling Mythos 5 for any foreign national. To comply, Anthropic disabled both models for every customer worldwide. Every other model in the index is unaffected.

Anthropic disagrees with the directive, calls it a likely misunderstanding over a narrow jailbreak it says is available in other models, and says it is working to restore access. There is no restoration date.

What this means for the index: Fable 5’s list price has not changed, but a model no one can currently buy does not meet our bar for stably available capability. We have marked it suspended and held it out, so the index returns to $2.17 across 19 members. When access is restored, Fable rejoins at $22 blended and the index follows. The full episode stays in the record.

This week’s issue asked whether anyone would pay for the most capable model on the market.

For now the question is blunter: whether it is one you are allowed to use.

On 9 June, Anthropic released Claude Fable 5, the first publicly available model from its Mythos class, the tier it positions above Opus in capability and had kept to a small set of vetted partners under its Project Glasswing program. By Anthropic’s own account, Fable 5 is more capable than anything the company has made generally available before. It launched at number one on the Artificial Analysis Intelligence Index, ahead of every other model the firm tracks, Anthropic’s own Opus 4.8 included.

It is also the most expensive model the index has ever tracked. Fable 5 costs $10 per million input tokens and $50 per million output, exactly double Claude Opus 4.8. At the index’s 70/30 blend that comes to $22 per million, comfortably above the previous ceiling, GPT-5.5 at $12.50.

This week it went into the index. Its subscription rollout is in flux (free until 22 June, then usage credits), but the API list price the index tracks, $10/$50, has been stable since launch. The index counts published list price, so Fable 5 qualifies as cleanly as any other member.

What it did to the index

Adding Fable 5 took the index from $2.17 to $2.43, a 12.3% rise in a week. Almost all of that move is one model, and it is worth being precise about what that means: the frontier did not reprice. The other nineteen members held flat. The index rose because a new, more expensive tier appeared above where the ceiling used to sit, and the index now reaches up to it.

That is the through-line this index keeps returning to. The market is not converging on a single price for frontier AI. It is stretching. The floor did not move this week; the ceiling nearly doubled. The gap between the cheapest model the index tracks, DeepSeek’s V4 Flash at $0.18 blended, and the most expensive widened from about 69 times to about 121 times in seven days.

The cheap tail keeps extending down while a higher rung keeps being added at the top.

The real question: will anyone pay for it?

A price of double Opus is a bet. Anthropic is testing whether a capability premium has buyers, probing the top of the willingness-to-pay curve rather than discovering a market-clearing price. The case for paying it is not really about tokens; it is about outcomes. If a more capable model finishes in one pass what a cheaper model needs five attempts to get right, $50 per million output can work out cheaper per finished task than a model a tenth of the price. Two dollars per million is not two dollars per outcome. That gap, between list price and realised cost, is the whole reason the simulator on the site exists.

Where that point sits is the interesting question for the next few months.

What I’m watching

1. Whether Fable 5’s price holds. The premium over Opus is the number to watch. If it drifts back toward Opus, or folds into standard plans at no surcharge once capacity allows, that is the clearest sign the market would not bear a capability-class premium.

2. Whether any other lab builds a rung above its flagship. If the premium were clearly working, you would expect OpenAI or Google to add their own above-flagship tier within a quarter or two. Silence at the top of their pricing ladders would be its own answer.

3. Whether premium prices keep climbing while usage moves down-tier. If list prices at the top rise while real workloads migrate to the cheaper, good-enough models, that is revealed preference: willingness to pay at the frontier is softer than the headline prices suggest.

Index data, simulator tools and methodology are at tokenpriceindex.com.

Bucking the trend

On 28 May, Anthropic released Claude Opus 4.8, its upgraded flagship. The price was the story, and the story was that there wasn’t one. That sounds unremarkable until you look at what the rest of the frontier did in the last few weeks:

• Alibaba launched Qwen 3.7 Max on 21 May at $2.50/$7.50, a number that would have looked steep for a Chinese model a year ago

• Days earlier, Google’s Fast tier slot in the index tripled when Gemini 3.5 Flash landed at $1.50/$9, replacing Gemini 3 Flash Preview at $0.50/$3.

• On 1 June, MiniMax replaced its own flagship with M3 at $0.60/$2.40 per million tokens. The model it succeeded cost $0.30/$1.20. A one-week launch promo briefly held the old rate, but the standard price is now double. MiniMax's doubling was what actually moved the index this week.

Recent flagships have often arrived more expensive than the models they replaced. The tier name stays the same. The price underneath goes up. Anthropic didn’t do that.

Opus 4.8 is positioned as an improvement on 4.7, and it shipped at the identical standard rate card. On Anthropic’s published pricing, its Fast mode, the option for lower latency, is markedly cheaper: $10/$50 against $30/$150 on 4.7. So the standard price held, the latency option price fell, and the capability rose. No quiet uplift, no repriced tier.

Why this is worth noticing

Frontier pricing has been moving in two directions at once, which is the throughline of this index. The cheap tail keeps getting cheaper. DeepSeek made a permanent cut only last month. The premium ceiling, meanwhile, has drifted up across several recent releases, usually through repricing at the moment of a model swap, when the higher number is easiest to wave through as a routine product update. Holding a flagship price flat across a capability upgrade runs against that current.

So what?

If you run workloads on a top-tier model, the version bump is a free upgrade rather than a budgeting event, which is rarer than it should be. It is worth re-running your evals on the new flagship before assuming the previous one is still your best option at that price point.

More broadly, the held price is a reminder that tier names are not prices. “Flagship,” “Pro,” and “Flash” carry different costs from one version to the next and from one provider to the next. The only way to know what you are paying is to track the rate card under the name, which is the entire reason this index exists.

What I’m watching

Whether Anthropic’s flat pricing holds at the next flagship release, or whether this was a one-off competitive hold. And whether any other lab follows by keeping a flagship price flat through an upgrade. Among the major frontier releases the index has tracked in recent weeks, Anthropic is the only provider that has.

The full historical series and methodology are at tokenpriceindex.com

Subscribe now

The bloc

Of the ten providers the Token Price Index tracks, five are now Chinese: DeepSeek, Alibaba, MiniMax, Moonshot, and Z.ai. Combined, they account for seven of the nineteen models the index follows. The cheapest of them cost a small fraction of the most expensive models in the index, while scoring close to them on the benchmarks most teams care about. A year ago that price gap came with a clear quality penalty. For a lot of everyday work, it no longer does.

A note on tokenizers

Tokens aren’t quite the same unit across models. Anthropic’s Claude family, for example, generates up to 35% more tokens for the same input than rivals, quietly raising its real cost above the rate card. That nudges the comparison. It doesn’t explain a gap that runs twentyfold.

Why the gap holds

These low prices are not a temporary land grab that snaps back once a marketing budget runs out. They are built into how the models are made.

The labs behind them design for efficiency, getting more output from less computing power. Limited access to leading-edge chips likely reinforced a focus on efficiency and inference optimisation. Several of these providers also choose to price low deliberately, to attract developers and build their ecosystems, earning their margin elsewhere. Low pricing is the strategy, not a promotion.

The twist

If cheap models were the whole story, the cost of frontier AI should be falling. It is doing the opposite, because the market is splitting in two and the expensive end is pulling harder.

The split widened quickly. In April, OpenAI launched GPT-5.5 at $12.50 per million tokens blended. The same month, DeepSeek launched V4 Flash at $0.18. The multiplier between the cheapest and most expensive model in the index jumped from 17x to 69x. It has held there since.

At the cheap end, the floor keeps extending. DeepSeek recently made a 75% cut to one of its flagship reasoning models permanent, dropping it to $0.57 blended per million tokens. It sits next to MiniMax’s flagship at the same price point, which has been there since March. Two frontier-class models at fast-tier prices is no longer an anomaly.

At the premium end, the ceiling keeps moving up. Google tripled the price of its fast tier last week. Alibaba’s newest flagship launched at a price that would have looked expensive a year ago. This week alone, the index added five new flagship-class models from five different providers, three of which replaced cheaper predecessors at higher prices.

So the floor is dropping while the ceiling rises. The index measures the blended cost of frontier capability, and that number is up 40% over the past month.

So what?

The distance between the two ends is now wide enough to change how you think about cost. A genuine choice exists that did not a year ago. Send simple, high-volume work to the cheaper models. Reserve the premium tier for the tasks that genuinely need it. Done well, that decision is no longer a rounding error on your bill. It is one of the largest levers you have.

What I’m watching

Whether other premium providers respond to DeepSeek’s permanent cut, or hold steady on brand. And whether flagship prices keep climbing across the board. If they do, the cheap-versus-premium divide will matter far more than where a model happens to be built.

The simulator on tokenpriceindex.com lets you model the split directly. Methodology and historical data are at tokenpriceindex.com/methodology.

Subscribe now

What Google said versus what the community said

Google’s framing was clean. Gemini 3.5 Flash beats Gemini 3.1 Pro on coding and agentic benchmarks at lower cost than Pro. Technically true, strategically chosen. Most teams using Flash tier weren’t picking Pro and downgrading. They were budgeting around the previous Flash price point. For them, the cost just tripled. The “cheaper than Pro” line is a bit like a supermarket raising the price of milk by 30%, then pointing out it’s still cheaper than oat milk.

Google’s own AI developer forum tells the other side. Thread titles in the days after launch include “Gemini 3.5 Flash is more expensive than 3.1?” and “Flash 3.5 is not a suitable replacement for Flash 3.0.” Multiple developers ran their existing Flash workloads against the new pricing and arrived at the same conclusion: the unit economics shifted materially against them.

Why this matters beyond Google

Frontier model pricing has been moving in two directions at once. New frontier capabilities launching at premium prices, and existing tier names being repriced upward when the underlying model is replaced. Product line names like “Flash” or “Mini” sound consistent across versions. The pricing underneath is not.

TPI as of 22 May 2026: $2.08, up 76.8% year-on-year from $1.18 in May 2025. Google’s Fast tier reprice contributed the largest single price increase this month.

Can developers just stay on the old model?

Technically yes. Gemini 3 Flash Preview remains live at its original pricing, with no deprecation date announced. Google’s official guidance is “strongly recommend updating” - pressure, not force. There’s even the option to remain on Gemini 3 Flash Preview.

But “preview” is doing real work in that name. Preview models are by definition not part of Google’s long-term support commitment. Production teams budgeting around a 3x cheaper preview model are taking on deprecation risk that doesn’t appear on the price page. The reprice isn’t predatory, it’s the upgrade path being made expensive while the alternative carries quiet risk. That’s a different kind of pressure, and arguably more interesting.

What I’m watching

Qwen 3.7 Max launched on 21 May at $2.50/$7.50, positioned as Alibaba’s flagship agentic model. It will likely join the index at this Friday’s sync. Chinese providers have historically priced aggressively below Western counterparts. $2.50 input is no longer aggressive - it sits alongside GPT-5.4 ($2.50) and approaches Claude Sonnet 4.6 ($3.00).

The frontier price floor may be rising globally.

The index updates weekly at tokenpriceindex.com - reply and let me know what you’re seeing in your own usage.

Subscribe now