Google lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Anthropic lists Fable 5.1 at $10 and $50. That gives Gemini a 13.33x standard-rate advantage on uncached input and output—a major procurement signal, but not proof that it produces the lower cost for a successfully completed coding task.
The distinction is more than accounting. Both platforms can consume tokens for internal reasoning, so the visible length of a coding answer does not reveal the full bill. Google includes thinking tokens in Gemini output pricing. Anthropic’s effort control applies to response text, tool calls and thinking, while Fable 5.1’s output tokens are separately priced.
The rate-card lead is simple—and time-limited
On standard, uncached usage through year-end, the comparison is direct. Gemini’s input rate is one-thirteenth of Fable’s listed rate, and its output rate is one-thirteenth as well. This is not a small optimization that can disappear in rounding; it is a wide gap before a workflow makes any other choices about how much context or reasoning to use.
Standard API prices through December 31, 2026
Input tokens
Gemini’s listed standard input rate is 13.33 times lower than Fable 5.1’s listed base input rate during the current pricing period.
Output tokens
Gemini’s listed output rate includes thinking tokens; the displayed rate gap is also 13.33 times.
The published Google schedule changes on January 1, 2027. Gemini’s standard input and output rates are set to double to $1.50 and $7.50 per million tokens. If Fable 5.1’s listed $10 and $50 rates were unchanged, the comparable gap would narrow to 6.67x—not remain 13.33x. Long-running deployments should therefore model both periods rather than treating today’s Google rate as permanent.
More reasoning can change both quality and cost
Anthropic presents effort as its main control for balancing intelligence, latency and token efficiency. Its levels run from low to max, with high as the default. Lower settings can reduce token use and speed responses, but Anthropic says they can also reduce capability. The company describes effort as a behavioral signal, not a strict token budget.
Google says Gemini 3.x models have thinking enabled by default, which can raise both latency and total token consumption. Developers can lower the thinking level or turn it off when cost or speed takes priority. A high-versus-high comparison may sound even-handed, but the vendors do not document like-named settings as equal amounts of inference work.
Caching rewards different request shapes
Repeated codebase context can substantially alter the bill. Google lists Gemini cached context at $0.075 per million tokens through December 31, plus $0.50 per million tokens per hour for storage. Anthropic separates cache creation from reuse: Fable 5.1 lists five-minute cache writes at $12.50 per million tokens, one-hour writes at $20, and cache reads at $0.25.
Anthropic’s caching system reuses matching prompt prefixes, making it most relevant when a system prompt, tools or long context remains stable across requests. That is a different economic pattern from a sequence of short, independent coding jobs. Cache pricing is therefore not a side detail: it can change which portion of a model’s rate card dominates a real workload.
What a matched coding-cost test needs
- The same coding tasks and a clear pass condition, including whether revisions count as another attempt.
- Configured reasoning and effort levels, because these settings can change token use, latency and capability.
- The same context and cache treatment, especially for codebase-heavy or multi-turn work.
- A stated retry policy. Google’s Python SDK automatically retries transient failures up to four times by default.
The success-rate hurdle is unusually high
The current price ratio establishes a useful conditional test. If both systems used identical uncached input and output volumes, with no other charges, Fable could cost less per successful completion only if its success rate exceeded Gemini’s by more than 13.33 times. Even perfect Fable completion could not overcome the rate gap if Gemini completed at least 7.5% of the same tasks.
Those thresholds are arithmetic, not measured coding performance. Anthropic advises customers to test Fable 5.1’s effort levels against their own evaluations, including a fresh sweep rather than carrying settings over from an earlier model. That is the sensible procurement response to the price gap: measure completed work under fixed settings and workflow rules, then compare the all-in cost.
Reader comments
Newest comments first. Replies stay oldest first.