AI API Pricing in 2026: Headline Rates Converged, the Metering Didn't
A single agentic coding request on GitHub Copilot burned 822 AI Credits, more than half of one Pro+ subscriber’s entire monthly allocation of 7,000. That happened three months after GitHub retired premium requests in favour of usage-based billing, and it is the sharpest illustration of what changed across AI pricing this year: headline rates look flat or falling, but what lands on the bill depends on metering details most teams have not read closely.
The frontier converged to within a few dollars
As of late August 2026, the three frontier API tiers sit close enough together that model choice is no longer a pricing decision on its own.
| Model | Input | Output | Notes |
|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | Anthropic’s frontier tier since Jul 24 |
| GPT-5.6 Sol | $4.00 | $20.00 | Promotional, list price is $5/$30 |
| Gemini 3.1 Pro (up to 200K) | $2.00 | $12.00 | Rises to $4/$18 above 200K input tokens |
OpenAI’s Sol cut, announced August 21, undercuts Opus 5 on both input and output, but only for the length of a promotion running through at least November 21. Its list price of $5/$30 is the most expensive of the three once the promotion lapses.
The one price move that is not promotional: Claude Sonnet 5 launched June 30 at $2/$10, with a scheduled September 1 increase to $3/$15 already cancelled. Anthropic’s own docs now state that increase “will not occur.” That is the rare case of a frontier vendor locking in a rate rather than testing how much the market will bear.
Every vendor now meters the same three axes
Base tokens are the number everyone quotes. The two numbers that move a production bill are cached input and batch processing, and all three vendors have converged on the same discount structure for both.
Cache reads cost roughly a tenth of standard input everywhere: $0.10 on Anthropic’s Haiku tier, $0.50 on Opus 5, and about 10% of the input rate on OpenAI and Google. Batch processing, for workloads that can tolerate asynchronous turnaround, is a flat 50% off input and output at all three vendors. Stack both together and effective spend on a cacheable, batchable workload can fall more than 90% below the headline rate, on any of the three platforms.
That is a bigger lever than choosing between Sonnet 5 and Gemini 3.1 Pro. A team that picks the “cheaper” model but never turns on caching is leaving more money on the table than the model choice was worth.
Consumer tiers cluster at $20, symmetric all the way to $200
The chat subscription market has converged even harder than the API market. ChatGPT Plus, Claude Pro, and Google AI Pro all price at approximately $20 a month, and the top individual tier on all three vendors lands at $200 a month: ChatGPT Pro’s 20x plan, Claude Max’s 20x plan, and Google AI Ultra’s top storage tier.
The one outlier is Google AI Plus at $4.99 a month, cut from $7.99 in June. Neither OpenAI nor Anthropic has a paid tier under $20, which makes Google AI Plus the only sub-$10 option with access to a full-capability model rather than a free-tier default.
Claude subscriptions do not include API access; that is billed separately at the rates above, a distinction worth confirming before assuming a Pro or Max seat covers programmatic usage.
GitHub Copilot swapped seats for a credit meter
On June 1, 2026, every GitHub Copilot plan moved to usage-based AI Credits, at 1 credit per US cent, replacing the old premium-request model. Code completions and next-edit suggestions stayed unlimited and free on every paid plan. Chat, agent mode, code review, and the CLI now draw from a credit pool billed at the underlying model’s per-token rate.
| Tier | Price | Credits included |
|---|---|---|
| Pro | $10/mo | $15 (1,500) |
| Pro+ | $39/mo | $70 (7,000) |
| Max | $100/mo | $200 (20,000), new signups paused |
| Business | $19/seat | $19 pooled per user |
| Enterprise | $39/seat | $39 pooled per user |
Overage bills at the same per-token rates once the pool runs out, and it is switched on by default for organisations unless an admin turns it off. The 822-credit anecdote from the opening is the kind of event that setup produces: one large agentic request, no cap in place, more than half a month’s Pro+ allocation gone in a single call. Teams whose usage is mostly inline completions will not notice the change at all, since completions never touch the credit pool.
What this means for anyone owning an AI budget
First, model choice matters less than the caching and batch decision. Route bulk or classification work to the cheapest current-generation model (Gemini 3.7 Flash or GPT-5.6 Luna sit at the current budget end), escalate hard reasoning and agentic coding to a frontier tier, and turn on caching and batch wherever latency allows on all of it.
Second, do not budget on promotional pricing. Sol’s $4/$20 and Gemini 3.7/3.6 Flash’s $0.75/$3.75 are both time-boxed, the OpenAI promotion through at least November 21 and the Gemini one through the end of the year, with Google’s own pricing page stating Flash reverts to $1.50/$7.50 on January 1, 2027. Model the reversion, not the promotion.
Third, if the team is on GitHub Copilot, set org-level or per-user spend caps before rollout rather than after. Overage is on by default, and a single agentic session can consume a meaningful fraction of a monthly allocation before anyone notices.
Fourth, revisit this comparison when any of three things happen: Google ships Gemini 3.5 Pro (it has been delayed indefinitely and would reset the frontier calculus), the OpenAI Sol promotion expires, or January 1, 2027 arrives and Gemini Flash pricing doubles. Any one of those changes the cheapest-path answer for workloads currently routed on today’s numbers.