Cheap tokens, expensive uncertainty
OpenAI’s latest GPT-5.6 pricing cuts and usage-limit fixes show how quickly AI model costs can fall while the underlying reasoning behavior and quotas remain unstable, forcing developers to treat model choice as a shifting operational risk rather than a static product decision. OpenAI introduced GPT-5.6 Sol as a more efficient, agentic coding model, with its CEO saying it was 54% more token efficient on agentic coding tasks compared with rivals while matching or beating their quality. At the same time, Terra and Luna launched as the balanced and low-cost tiers for everyday and high-volume work. On paper, this looks like a simple menu: pick Sol for ambitious agents, Terra for daily workflows, Luna for scale. In practice, the last few weeks have made that decision far less straightforward.
GPT-5.6 pricing cuts: aggressive move, clear signal
OpenAI has pulled a classic platform move: slash prices hard, then backfill the margin gap with efficiency gains. GPT-5.6 Luna, the fastest and cheapest tier, now costs 80% less than its initial price, while Terra is 20% cheaper starting July 30. The new API rates are USD 2 (approx. RM9.20) per million input tokens and USD 12 (approx. RM55.20) per million output tokens for Terra, and USD 0.20 (approx. RM0.92) per million input tokens and USD 1.20 (approx. RM5.52) per million output tokens for Luna. Pricing for GPT-5.6 Sol stays at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens. OpenAI attributes the cuts to efficiency wins across its stack, including kernel rewrites that reduced serving costs by 20% and token-generation experiments that lifted efficiency by more than 15%. This is not defensive discounting; it is a statement that OpenAI wants to own the cost-per-task curve.
The pricing changes also flow through to subscription users. Because Terra and Luna now consume fewer credits, ChatGPT Work and Codex subscribers can run more tasks before hitting quota. For startups obsessed with developer API costs, the message is seductive: you can get performance comparable to last year’s frontier-class models at a fraction of previous prices, with OpenAI claiming Luna beats a major rival’s model on professional work at an estimated cost per task nearly 99% lower. The catch is that low cost is only one half of reliability. If model behavior and metering keep moving, a cheap platform can still be an expensive dependency.
Usage limits, tool use, and the hidden cost of Sol
While Luna and Terra became cheaper, Sol revealed how costly powerful agents can be when their behavior surprises even the vendor. Shortly after launch, Codex and ChatGPT Work users saw their usage vanish during long agentic coding sessions, where Sol continued working while tools ran in the background. In an X post, engineering lead Thibault Sottiaux admitted that Sol’s new programmatic tool calling, or “code mode,” meant the model coordinated more complex workflows and made more tool calls than expected, driving far higher token use than OpenAI’s tests predicted. The company conceded it had focused on average and median usage before launch and “missed some cases where the long tail could use significantly more usage”.
That miscalculation had direct consequences: power users saw their Codex allowances burn down far faster than they could reasonably forecast. OpenAI’s fix was decisive but reactive—reset usage limits for all ChatGPT Work and Codex users and roll out backend inference improvements that should make typical Sol sessions about 18% longer before hitting limits. Another set of optimizations earlier promised around 10% more usage for Sol users as well. The update solves an immediate pain point, but it exposes the larger issue: once agents are free to run, parallelize, and retry out of sight, predicting how much of a subscription allowance any single job consumes becomes much harder for developers.
Reasoning limits and silent model changes
More worrying than usage resets is how Sol’s reasoning behavior shifted under production workloads. OpenAI did change GPT-5.6 Sol’s internal reasoning settings after launch, then rolled those changes back when users noticed. Public discussion focused on “reasoning effort” or “juice” values—internal knobs that govern how much compute the model spends thinking through a problem. One widely shared claim that the Max-tier reasoning budget dropped from 960 to 128 overnight could not be verified and has been walked back, but OpenAI itself acknowledges it reverted experiments that altered reasoning efforts. At the same time, it was fixing heavier-than-expected multi-agent behavior at high and xhigh settings and had to roll the Codex context limit back from 372,000 to 272,000 tokens after “higher-than-intended usage charging”.
This is where trust gets tested. Production users experienced changing behavior—different reasoning budgets, different context handling, different multi-agent dynamics—before receiving a clear explanation. If you priced your own product around Sol’s launch behavior, and then OpenAI changed these invisible settings under load, your margins could move even if your product stayed exactly the same. AI model reasoning limits are no longer an academic detail; they define latency, reliability, and cost. Treating them as invisible implementation details is untenable for anyone building a real business on top of the OpenAI API changes.
What this means for developers betting on GPT-5.6
The pattern across the last few weeks is clear: OpenAI can cut GPT-5.6 pricing by up to 80%, improve token efficiency, and even replace Priority Processing with a Fast mode that delivers Sol responses up to 2.5× faster at twice the price with no change in intelligence. But it has also shown a willingness to tweak reasoning behavior, context limits, and usage metering on live systems without the kind of advance notice infrastructure buyers usually expect. GPT-5.6 Terra and Luna remain available through the API and subscription products, and on a cost-per-token basis the value proposition is strong. The open question is whether startups can rely on the models’ behavior staying stable enough to build precise pricing and SLAs on top.
For developers, the takeaway is uncomfortable but useful: model capability is now an operational dependency, not a fixed product description. The Sol episode does not prove malicious intent; it shows how tightly cost optimization and model capability are intertwined. If you are evaluating long-term API dependencies, you should treat GPT-5.6 pricing cuts as one input, and demand more explicit guarantees on reasoning limits, context windows, and metering rules as another. The future will belong to teams that treat AI models like any other critical infrastructure: benchmark them under real workloads, monitor them continuously, and negotiate for clarity before betting the business on them.
