Google's Standard API prices double on December 1. A 657-task automation benchmark already shows the bigger buyer decision: medium reasoning effort nearly matches high on success rate at much lower cost.
Google has opened Gemini 4 Carbon through its paid Gemini API tier and AI Studio, with Standard input and output token prices scheduled to double on December 1, 2026. The sharper commercial question is how much extra successful work customers get when they pay for more reasoning. Whether Google can charge more per token matters less.
Zapier's 657-task AutomationBench gives a useful, if limited, answer. Carbon's medium thinking setting passed 53.0% of tasks, against 53.9% at high. CTOL's estimated introductory API spending was $0.742 per attempt at medium and $0.9708 at high. Across the benchmark, that works out at approximately $1.40 versus $1.80 spent per recorded success.
For enterprise teams running agent automation similar to this benchmark, medium is therefore the stronger starting point. High costs approximately 31% more per attempt for a measured gain of only 0.9 percentage points in task success. The finding does not establish the economics of every production workload. It does give buyers a concrete reason to demand task-specific evidence before choosing the most expensive reasoning setting.
Medium nearly matches high at lower cost
Carbon supports low, medium and high thinking levels, with medium the default. Google says higher settings generally increase token consumption and latency, and thinking tokens are included in output-side billing.
The three settings produced substantially different results on Zapier's AutomationBench.
| Thinking level | Task pass rate | Estimated cost per attempt through Nov 30 | Estimated cost per attempt from Dec 1 | Estimated spending per recorded pass, Nov → Dec |
|---|---|---|---|---|
| Low | 31.5% | $0.4271 | $0.8542 | $1.36 → $2.71 |
| Medium | 53.0% | $0.7420 | $1.4840 | $1.40 → $2.80 |
| High | 53.9% | $0.9708 | $1.9415 | $1.80 → $3.60 |
Source: Zapier AutomationBench pass rates and CTOL-estimated costs using Google's Standard API tariffs. The pass rates are observed benchmark outcomes; the December costs are tariff sensitivities, not observed invoices. Spending per recorded pass equals estimated spending per attempt divided by the corresponding pass rate. The underlying token counts for CTOL's per-attempt cost estimates are not itemized here, so the derived dollar comparisons remain conditional.
The marginal return from additional thinking is the comparison that matters.
Moving from low to medium raised the pass rate by 21.5 percentage points and increased estimated spending per attempt by about $0.315. Moving from medium to high added another $0.229 per attempt but delivered only 0.9 percentage points of additional success.
The low setting had the lowest estimated spending per recorded pass, at approximately $1.36. Medium was only slightly higher, at $1.40, while producing substantially more successful task outcomes. That makes medium an attractive first setting to test when task completion is important and failures impose operational costs.
High performed differently. Its estimated spending per recorded pass was approximately $1.80, nearly 29% above medium.
At the chart's estimated cost ratio, high would need a pass rate of approximately 69.3% to match medium's spending per recorded success. Its measured result was 53.9%.
The 69.3% threshold is no forecast of what Carbon could achieve on another task mix. It is the break-even success rate implied by the benchmark's estimated costs and medium's observed performance.
It also shows why choosing the highest reasoning setting by default is a poor procurement rule. Additional reasoning has value when it improves completed work enough to justify the extra expenditure, and the model doing more computation is not enough.
December doubles the dollar cost and leaves the ranking alone
Through November 30, Google's published Standard prices are $2 per million input tokens and $10 per million output tokens, including thinking. From December 1, the listed prices rise to $4 and $20.
Cache reads also increase from $0.10 to $0.20 per million tokens. Google says implicit caching is enabled and explicit caching is not supported.
At unchanged token consumption, those increases double the modeled Standard API cost of each effort setting. The ratio of high-effort spending to medium-effort spending stays the same.
The implications differ by workload.
For a fixed budget, higher absolute spending can limit how many tasks a team can run. For a workflow that generates measurable economic value from each completed task, buyers need to compare the higher API expenditure with that value and with the costs of failure, retries and human review.
The benchmark makes the price sensitivity visible but does not supply those production variables. Its spending-per-pass measure combines attempted-task costs with recorded passes, so it cannot predict the full cost of operating a business process.
The December increase therefore strengthens the case for task-level cost controls. By itself it does not establish that customers will abandon Carbon or accept the new tariffs.
Carbon must justify its extra reasoning against Argon
Google positions Carbon as its strongest Gemini 4 reasoning model for complex software engineering, scientific reasoning, cybersecurity and long-horizon agents.
The product offers a unified 1,048,576-token context budget for input and output, with a synchronous output limit of 65,536 tokens and a separately documented long-decode path. It accepts text, image, video, audio and PDF inputs. It also supports code execution, computer use, function calling, grounding and other agent-oriented tools.
Those capabilities broaden the workloads Carbon can attempt. They do not guarantee successful completion or economical operation.
Google's migration documentation gives buyers a particularly relevant comparison. Carbon and Argon share the advertised million-token context, built-in tools and per-token prices. Google also says Carbon spends more thinking tokens per turn and is expected to have higher latency and cost per turn.
Identical token tariffs therefore do not mean identical costs for a completed task.
An existing Argon customer should compare task success, total billed tokens and execution time before replacing it with Carbon. The supplied evidence includes no comparable Argon success rate for Zapier's task set. It cannot establish whether Carbon's additional reasoning overhead is justified against that incumbent.
Google's model card reports engineering, terminal, computer-use and cybersecurity evaluation results for Carbon. Those results support its positioning across demanding technical tasks. They do not substitute for a matched cost-and-completion comparison.
Paid access makes the decision actionable
Carbon is excluded from the Gemini API Free tier but available through paid API access and AI Studio. Google's model card also lists distribution through Gemini App, Gemini for business, Google Antigravity, Google AI Mode and the Fairwind Program for verified cybersecurity defenders.
For enterprise buyers, Google lists higher rate limits, Batch API access and caching capabilities on its paid tier, alongside a stated policy that content is not used to improve products. Enterprise offerings add dedicated support, security and compliance capabilities, provisioned throughput and volume discounts.
These are relevant operating and procurement features. They make it possible to evaluate Carbon within a managed commercial deployment. They do not show how widely customers are using it.
The immediate deployment question is narrower and more useful: which setting meets the buyer's required completion quality at the lowest total operating cost?
For tasks resembling the Zapier benchmark, the reported evidence favors testing medium before high. For work where an unsuccessful first attempt is inexpensive, low may be sufficient. High becomes attractive when a buyer can demonstrate materially better task outcomes on the workload that matters, including any avoided retries, manual intervention or failure costs.
What would change the view
CTOL's current assessment, with moderate confidence, is that medium reasoning effort is the more economical starting point for Carbon automation workloads resembling Zapier's benchmark. The evidence is the near-identical medium and high pass rates, the substantial per-attempt spending premium at high, and the resulting difference in estimated spending per recorded success.
December 1, 2026, is the next fixed pricing observation. If Google's documented Standard tariffs take effect, the modeled dollar costs double for unchanged workloads. That event will confirm a change in published pricing. It will say nothing about buyer acceptance or production adoption.
A comparable task evaluation conducted on or after December 1 would provide a stronger test of the deployment decision. At the same estimated high-to-medium cost ratio, a high-effort pass rate around 69.3% would eliminate medium's measured cost-per-success advantage on this benchmark. A buyer's own repeatable completion and spending data could also overturn the recommendation for its particular workload.
Until such evidence emerges, the supported decision is to benchmark medium first and make high prove its premium. Carbon's commercial value will depend on completed work per dollar. The maximum amount of reasoning it can perform does not decide that.