On July 30, OpenAI cut GPT-5.6 Luna's API price by 80% - from $1 to $0.20 per million input tokens, $6 to $1.20 per million output tokens. Starting July 30, API pricing dropped to $0.20 per million input tokens and $1.20 per million output tokens for Luna. That is a significant cut three weeks after the model launched. OpenAI announced the reduction roughly three weeks after GPT-5.6's public release, facing pressure to cater to a more cost-sensitive customer base where enterprises have been less inclined to deploy expensive models without a clear return on investment.
You would expect enterprise AI bills to be falling. They are not.
Blended AI token prices fell roughly 67% year over year - from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026. By every classical reading of a market, AI should be getting cheaper. Yet 73% of enterprises exceeded their original AI cost projections last fiscal year. The contradiction has a precise cause, and it is not waste or inefficiency. It is architecture.
Why agentic AI inference cost doesn't follow the token price
Agentic AI inference cost is not chatbot inference cost with a nicer interface on top. It is a structurally different workload.
A simple chatbot query triggers one inference call, but an agentic workflow - where a model calls external tools, verifies outputs, and self-corrects - can trigger 10 to 20 model calls for a single user-initiated task. That changes the relevant unit: it is no longer cost per prompt, but cost per completed task.
Gartner's March 2026 analysis puts agentic workloads at 5 to 30 times more tokens per task than a standard chatbot, and typical production agents land between ten and twenty model calls for a single user request.
The compounding is structural, not a configuration mistake:
Context growth: Each tool call appends results to the running context. By step five, the model is reading everything that happened in steps one through four.
RAG overhead: RAG architectures inflate context windows 3-5x on top of the base prompt.
Always-on agents: Always-on monitoring agents scan logs, inboxes, and market data around the clock, consuming compute whether or not a human is watching. These background workloads barely existed in enterprise budgets two years ago. Today they represent a growing share of inference spend that most finance teams never approved, because nobody itemized it.
EY's analysis found the cost of a single agentic customer-service interaction rose from roughly $0.04 in 2023 to $1.20 in 2026, a roughly 30-fold increase driven by orchestrated multi-tool workflows replacing linear chatbot exchanges. That is the real comparison to make against a falling token price: the per-token rate dropped, the per-interaction rate climbed.
The GPT-5.6 pricing structure is worth reading carefully
Luna is not one fixed capability number. It has an inference-time effort curve. The model can answer quickly with no reasoning, or spend far more tokens exploring, checking, and revising before it returns an answer. The quality gain can be substantial, but so can the effective cost and latency.
This matters for anyone routing agentic workloads to Luna on the assumption that $0.20 input is the whole story. The model supports a context window of 1,050,000 tokens, a maximum input of 922,000 tokens, and up to 128,000 output tokens. The pricing page also lists $0.40 input and $1.80 output for long-context Luna requests, with the higher rate applying to the full request.
If anything in your stack hardcodes the old prices, your cost numbers for Luna are now inflated five times. Price cuts sound like pure good news, and mostly they are. But every price cut also silently breaks every hardcoded price table that references the old numbers: budget alerts, routing rules, margin dashboards, and cost estimates.
The GPT-5.6 tier structure also clarifies what each model is actually for:
| Model | Input (standard) | Output (standard) | Best fit |
|---|---|---|---|
| GPT-5.6 Luna | $0.20/M | $1.20/M | High-throughput, low-latency: summarization, classification, routing |
| GPT-5.6 Terra | $2.00/M | $12.00/M | General production; balance of capability and cost |
| GPT-5.6 Sol | $5.00/M | $30.00/M | Complex reasoning, multi-step planning, tool-heavy agents |
Sol is aimed at the most complex reasoning-heavy and agentic workloads, including advanced coding, multi-step planning and tool-using systems. Luna is positioned for high-throughput, low-latency tasks such as summarization, classification, routing, and lightweight real-time assistants where cost per request is the primary constraint.
The practical takeaway: route planning and tool-calling steps to Sol, route output formatting and classification steps to Luna. A single-model agentic pipeline that runs everything through Sol is paying $30/M output for work that Luna can do at $1.20/M.
What the no-budget-cap problem costs in practice
No major agent framework ships with a native dollar-denominated budget cap. LangChain, CrewAI, and AutoGen all expose token counting as a post-hoc observable, not a pre-execution policy. Teams discover they've spent $47,000 after the fact.
Uber's CTO reported burning the company's entire 2026 AI coding budget in four months. Engineers liked the tools, agent modes consumed more tokens per task than the budget assumed, and procurement had no easy lever to throttle usage without breaking workflows. That is not a story about reckless spending. It is a story about the default state of most agentic deployments: visible output, invisible meter.
The three places to intervene, in order of ROI:
- Model routing - match model tier to step complexity. Planning and tool selection call the heavy model; formatting and verification do not.
- Context discipline - summarize earlier turns before feeding them back into a long loop. A 50K-token context that could be a 10K summary is burning 5x on every subsequent step.
- Semantic caching - semantic caching returns 30-60% savings in high-repetition scenarios. For an internal knowledge assistant answering the same 50 questions in different phrasings, caching is the cheapest optimization available.
A teammate like Beagle, which runs inside Slack, is a useful reference point for what context discipline looks like in practice: every draft cites the source it read, so the model does not re-read the same document on the next turn.
Beagle in action#product, 2:38pmAgentic AI inference cost: common questions
Why is my AI bill rising when token prices keep falling?
Per-token prices have fallen roughly 67% year over year, but agentic workloads consume 5-30 times more tokens per task than a chatbot, per Gartner's March 2026 analysis. A workflow triggering 10-20 model calls compounds that multiplier. Usage volume has outpaced price reduction - the result is higher absolute spend despite lower unit cost.
What is the right way to model agentic inference cost before deployment?
Model per workflow step, not per user request. Count the expected number of LLM calls per task, estimate the context window at each step (accounting for accumulating tool results), and multiply by your output token rate. Compare that per-task cost against the business value of automation - not against a chatbot cost estimate.
Does GPT-5.6 Luna's 80% price cut change agent economics?
It changes cost on the correct workloads: summarization, classification, routing, and output formatting. It does not change the cost of reasoning-heavy planning steps, which still belong on a higher-tier model. The full saving is only captured if you split steps by complexity rather than running a single-model pipeline.
Should every agent step use the cheapest available model?
No. Cheaper models trade cost for reliability on complex steps. A planning step routed to an under-powered model that mis-reasons and retries twice ends up more expensive than routing it to the right model once. The correct question is which step in this loop is reliably solvable by a smaller model - not which model is cheapest overall.
What does 'cost per completed task' mean in practice?
It means measuring inference spend normalized to one unit of business output - one support ticket resolved, one PR reviewed, one draft approved - rather than one token or one API call. That framing exposes whether automation is actually cheaper than the manual alternative, and it is the unit that survives the agentic multiplier intact.