Meta opened its first paid model API on July 9 and priced Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens. Those rates are below the flagship offerings from OpenAI and Anthropic, though the size of the discount varies by model and by token type. For an enterprise buyer, the rate card is only one term in the bill.
Agentic systems turn a single user request into repeated rounds of planning, retrieval, tool calls, validation and retries. The price per token can therefore fall while the cost per completed task, or the organization’s total AI spending, rises. That outcome is not yet established as a universal enterprise trend, and public data on production agent costs remains thin. The budgeting lesson holds even so, which is that finance teams should model successful workflow outcomes rather than multiply expected calls by the advertised price per million tokens.
A Price War Accelerates
Three near-frontier launches landed inside eight days. SpaceXAI shipped Grok 4.5 on July 8. OpenAI made the GPT-5.6 family generally available on July 9, and Meta opened the Meta Model API the same day. DeepSeek belongs to the same story from a distance, having made a 75% cut to V4-Pro pricing permanent in late May.
OpenAI published rates of $5 input and $30 output per million tokens for the flagship Sol, half that for Terra, and $1 and $6 for Luna. Those rates are current as of July 13. All four announcements placed unusual emphasis on price and performance per dollar, even as they went on making capability claims about coding, computer use and multi-agent orchestration. OpenAI told developers it had trained GPT-5.6 to get more useful work from every token.
Apart from the rate cards, the launches are based on an assumption borrowed from cloud. Unit prices fall every year, so inference should eventually become a rounding error in the operating budget. Storage and compute prices did fall for a decade. Total cloud bills still rose, as consumption, managed services and egress expanded. Agents may repeat that pattern on a shorter clock, because one user-visible action can trigger dozens of metered operations.
Four Numbers That Are Not The Same Number
Most agent budgeting arguments collapse because they treat four distinct quantities as one. Price per token is what appears on the rate card. Tokens per attempt is what the workflow actually draws. Cost per successful task divides the whole workflow cost by the tasks that finished correctly. Total organizational spending multiplies all of that by task volume.
These move independently, and sometimes in opposite directions. A more expensive model that resolves 90% of tickets on the first attempt can cost less per resolved ticket than a cheap model that resolves 40%, retries, escalates and eventually lands on a human. Total spending can rise even when cost per completed task falls, simply because more teams are running more agents.
The identity worth modeling is simple enough. Total spend equals task volume, multiplied by attempts per task, multiplied by tokens per attempt, multiplied by effective token price, plus tool and infrastructure costs. Only the fourth term in that chain is falling.
Why Agent Workloads Draw So Many Tokens
A conventional chat interaction usually needs one principal generation, even when retrieval, moderation and routing sit around it. An agentic workflow adds planning, tool calls, validation and retry loops on top, and it resends context every time around. The user sees one answer and the vendor pays for the whole loop.
Goldman Sachs forecasts that token consumption will multiply 24 times between 2026 and 2030, reaching 120 quadrillion tokens per month. The bank attributes the surge to always-on enterprise agents rather than to more people asking more questions. That is a market forecast, not a per-task measurement. Studies of agentic coding report per-task usage orders of magnitude above simple chat. The same studies report wide variance between repeated runs of an identical task, so no single universal multiplier holds across workloads.
The arithmetic still bites the budget. Illustratively, if the token draw for a task rises 20-fold while the unit price falls 75%, total model charges rise fivefold. Actual results depend on retries, caching, routing and tool fees.
Reasoning modes push in the same direction. Muse Spark 1.1 bills its internal thinking tokens at the output rate, so a heavy chain-of-thought call costs like a long answer. GPT-5.6 offers an ultra setting that coordinates four agents in parallel, which trades token draw for speed.
Where The Money Actually Goes
Whether input or output dominates the bill depends on the workflow. Agentic coding generates enormous input volume, because the system resends repository context, tool results and conversation history on every turn. Research on agentic coding finds input tokens driving the total. Workloads with long generated documents or heavy reasoning tilt the other way.
Take 10 million uncached input tokens and 1 million output tokens in a month. On Muse Spark 1.1 that is $12.50 of input and $4.25 of output. On GPT-5.6 Sol it is $50 and $30. Input is the larger line item in both cases, despite output carrying the higher rate. A team that reasons from output prices alone will size its budget against the wrong half of the bill.
Caching is the lever that changes those numbers most, and the terms differ enough between providers to matter architecturally. Meta reads cached input at $0.15 per million tokens. OpenAI keeps a 90% discount on cache reads for GPT-5.6 and bills cache writes at 1.25 times the uncached input rate. Anthropic charges separately for cache writes, at 1.25 times input for the five-minute window and twice input for the one-hour window. Google can add a per-hour storage charge on explicit context caching. Provider-specific cache breakpoints, retention windows and storage charges can create switching costs when an application is tuned tightly around one vendor’s caching behavior, although prompt abstraction layers reduce that exposure.
The Challenges
Token charges are not the agent bill. A production workflow also pays for search and retrieval, vector storage, reranking, browser and computer-use sessions, sandboxed code execution, observability, and the human review that catches what the agent got wrong. Any cost model that stops at the model API is measuring a fraction of the invoice.
The evaluation layer is weaker than the pricing debate assumes. OpenAI audited SWE-Bench Pro, estimated that roughly 30% of its tasks were broken, and said it no longer recommends the benchmark as a leading coding eval. Buyers who wanted a public score to certify that a cheaper model is good enough now have to run a workflow-specific evaluation instead, and that evaluation costs real money.
The negotiating leverage is also conditional. Public price competition helps buyers only where models are substitutable. Where quality differs materially, where a workflow depends on proprietary tools, where volume commitments are already signed, or where compliance narrows the provider list, a lower list price changes very little.
In summary, the rate card is falling and the invoice is a separate question. Enterprises should budget per successful outcome, instrument the loop before scaling it, and date every price they model, because these numbers moved three times in eight days.
Three questions belong in the next vendor conversation. What is the median token draw per successfully resolved task, counting retries and escalations? Which steps in the workflow can run on a Luna-tier or Muse-tier model without a measurable quality regression? What happens to the bill when tool calls double while each metered layer bills independently?
The price war gives enterprise customers a stronger position than they have had since inference became a line item. The customers who convert that into leverage will be the ones who walk in with their own cost-per-outcome numbers rather than the vendor’s rate card.
