Gemini 4 vs Claude Fable 5.1: Add an Agent Route or Hold?
Status and Evidence Table
| Question | Gemini 4 | Claude Fable 5.1 | Safe decision |
|---|---|---|---|
| Official status | Pre-training confirmed by Google (July 2026) | Released by Anthropic on September 1, 2026; marked Active (latest) | Treat the gap as a status gap, not a capability verdict |
| API model ID | Not published | claude-fable-5-1 (Anthropic) | Never place an unpublished ID in configuration or code |
| Context and max output | Not published | 1M context; 128K max output (Anthropic) | Record how much of the window your traces actually consume |
| Input modalities | Not published | Text and image input, text output (Anthropic) | List the modalities each workload requires before testing |
| Thinking controls | Not published | Adaptive thinking always on; default effort high; comparative latency listed as Slower (Anthropic) | Log the effort setting per trace so a later route is tested at matched settings |
| Cache pricing structure | Not published | Input $10 / output $50 per MTok; 5-minute cache write $12.50; 1-hour cache write $20; cache read $0.25 per MTok (Anthropic list) | Cache economics cannot be compared until both sides publish TTLs and rates |
| Batch availability | Not published | Anthropic lists Batch input and output at 50% of standard | Compare Standard with Standard, Batch with Batch |
| Documented breaking changes | Not published | Three in Anthropic's migration guide: forced tool choice errors, older-model incompatibility with 5.1 thinking blocks, invalidated thinking blocks after earlier-turn edits | A published breaking-change list is testable; an absent one is not |
| EvoLink route | No verified callable route; launch alert open | Configured route claude-fable-5-1; pricing on the product page | Verify returned model identity on the first request of any route |
The Gemini 4 column is a list of what can be measured, not a criticism. Every "Not published" cell is a field your harness should already have a slot for, so a published contract fills in values instead of forcing a redesign.
Which Long-Running Tasks Are Worth Evaluating
Long-running agent work is not one workload. Five task classes cover most traces teams route to Fable 5.1, each with its own definition of "accepted". Record the baseline per class; judge any future Gemini 4 route per class as well.
| Workload | What "accepted" means | What to record on Fable 5.1 now | What a Gemini 4 route must later match |
|---|---|---|---|
| Multi-file repository repair | Tests pass, diff reviewed, no unrelated files touched | Files read and edited, tool calls, retries, reviewer minutes, tokens per accepted fix | Same task set, test suite and reviewer rubric |
| Multi-step research and knowledge deliverables | Citations resolve, structure matches brief, no unsupported claims | Sources fetched, revisions requested, factual corrections per deliverable | Same briefs and fact-check pass |
| Multi-hour tool loops | Loop ends on the stop condition with a valid final state | Loop count, failed tool calls, checkpoints, resume events, wall-clock to termination | Same tools, permissions, retry budget and stop rules |
| Structured extraction over long documents | Output validates against schema; sampled fields match source | Document length, schema failures, re-runs, field accuracy on a fixed sample | Same documents, schema and sample |
| Customer-service multi-step tool flows | Ticket resolved inside policy without human escalation | Escalations, wrong-tool calls, policy violations, handle time | Same flow definitions and policy checks |
No row names a preferred model, because no matched run exists. The table only fixes what "the same task" will mean when a second route becomes available.
Context Reuse and Cache Conditions
Long-running agents re-send the same repository map, tool schemas and instructions on every turn, so cache pricing dominates the token bill. It is also the part of the comparison that cannot start today.
Google has not published Gemini 4 cache TTLs, write or read rates, cache scope (automatic prefix caching versus explicit cache objects), or a minimum cacheable size. Without those four values there is no denominator: a five-minute cache read is a different product from a one-hour read, and a per-request prefix cache is a different product from a named cache several agents share. Different TTLs, context scopes and batch modes cannot be mixed in one number.
Fable 5.1 gives you the structure to record against. Anthropic documents a 5-minute cache write at $12.50 per MTok, a 1-hour write at $20 per MTok, cache reads at $0.25 per MTok, and Batch at 50% of standard input and output. For every Fable trace, log:
- which TTL was chosen and why (turn spacing under five minutes, or gaps that justified the 1-hour write);
- cached prefix size and how often it changed mid-task, since a changed prefix forces a new write;
- cache-read tokens versus fresh input tokens per turn;
- whether the trace ran Standard or Batch, and how Anthropic's pricing page combined Batch with cache modifiers at the time of the run.
Retry, Recovery and Human Takeover Cost
This is the layer a spec chart never shows and where most long-running agent budgets are actually lost. A trace that completes on the third attempt after a human unblocked it does not cost the same as a trace that completed once. Log six events per trace, each with a timestamp and token count:
- Loop detection. The same tool call with the same arguments repeated past a threshold set in advance. Record how many turns burned before the stop rule fired.
- Failed tool calls. Separate provider-side errors, tool-side errors and model-side malformed calls. Only the third is a model quality signal; the other two are infrastructure cost a new route will also pay.
- Checkpoint integrity. Verify each saved state can actually restore the task. A checkpoint that cannot resume is wasted output plus a full re-run.
- Resume after interruption. Count the tokens spent rebuilding context after a rate limit, timeout or operator pause, separately from first-run tokens.
- Thinking-block invalidation. Anthropic's migration guide documents that editing earlier turns invalidates retained Fable 5.1 thinking blocks. Any framework that rewrites history for compaction or correction will trigger this; record how often, and what the re-think costs. Whether Gemini 4 has an equivalent rule is not published.
- Human takeover. Reviewer minutes spent unblocking, correcting or finishing the task, logged as time so it never disappears inside an API total.
Store these as fields on the trace record, not in a separate incident log; cost per accepted task is only meaningful if retries, resumes and reviewer minutes sit on the same task ID as the tokens.
The Full Task Cost Variable Table
The denominator for every figure below is tasks that completed and passed acceptance. Tasks that completed but failed review count in spend (they were paid for) but not in the denominator. API spend and human time stay in separate columns; convert minutes to money only at the reporting layer, at a rate your finance team owns.
| Variable | How to record on Fable 5.1 today | Gemini 4 status |
|---|---|---|
| Input tokens | Fresh (uncached) input per turn, summed per task, from the usage counters EvoLink returns | Unknown until contract published |
| Cache write | Tokens written, split by 5-minute and 1-hour TTL, with the write count per task | Unknown until contract published |
| Cache read | Tokens served from cache per task; also log hit ratio (cache read / (cache read + fresh input)) | Unknown until contract published |
| Output tokens | Final answer plus intermediate tool arguments, summed per task | Unknown until contract published |
| Tool calls and external cost | Count and any metered cost (search, code execution, third-party APIs) per task | Unknown until contract published |
| Retries | Attempts beyond the first, with tokens per attempt, attached to the same task ID | Unknown until contract published |
| Fallback route usage | Whether any turn was served by a different model, and which one, verified by the returned model field | Unknown until contract published |
| Human review minutes | Reviewer and operator time per task, recorded as minutes | Measured on your own runs once a route exists; not a vendor-published figure |
| Wall-clock to acceptance | Time from task start to accepted result, including waits and human turns | Measured on your own runs once a route exists; not a vendor-published figure |
| Accepted-task rate | Accepted tasks / attempted tasks per workload class over the window | Measured on your own runs once a route exists; not a vendor-published figure |
Cost per accepted task is API spend divided by accepted tasks, reported next to review minutes per accepted task. A route that lowers the first while raising the second has not reduced cost; it has moved cost onto people.
Completion Quality Acceptance
An acceptance rubric that only works for one model is a preference, not a rubric. Build acceptance so it survives a model change:
- Objective checks first. Tests pass, schema validates, citations resolve, the diff touches only declared files. Binary and cheap to run on every trace.
- Reviewer rubric second. A short fixed checklist per workload class (correctness, completeness, safety, adherence to brief), scored by reviewers who do not know which route produced the output.
- Repeated runs. Run each task more than once and report the accepted-task rate as a range, so a one-run success is not promoted as a capability.
- Uncertainty reporting. Publish sample size, settings (effort, tools, TTL) and the failure list next to every result. A rate without a sample size is not evidence.
Write the rubric before any Gemini 4 access exists; written afterwards, it will bend around whatever the new route happens to do well.
Shadow Traffic and Rollback Plan
When a Gemini 4 route eventually exists, it should enter through the same three stages as any new route, with Fable 5.1 retained as the measured baseline throughout.
- Replay. Feed recorded Fable 5.1 traces to the candidate offline; compare accepted-task rate and cost per accepted task on identical inputs.
- Shadow. Send live requests to both routes, serve the Fable 5.1 answer, score the candidate silently. This surfaces loops and tool errors without user exposure.
- Canary. Route one workload class at a small share; expand only when gates hold across the observation window.
Promotion gates, set before the test starts: accepted-task rate at or above baseline; p95 wall-clock to acceptance within the agreed bound; cost per accepted task, including retries and fallback, at or below baseline; zero incidents for the workload class over the window.
Rollback trigger: any gate failing for the agreed number of consecutive days, or one incident that reaches a customer. Rollback is a gateway configuration change, not a code deploy, which is why model selection belongs in routing config rather than application code.
model field and usage counters in every response and checking they match the configured route. On a gateway, this confirms a request was served by the model you asked for and not silently by a fallback. Do it on the first request of any new route and continuously in shadow; record the result in the fallback-route field of the cost table.What Would Not Count as Progress
- A larger rumored Gemini 4 context window with no completed long tasks to show for it.
- A lower headline price that is offset by retries, resumes, longer outputs or extra review minutes.
- Leaked benchmark screenshots or third-party spec tables without settings, sample size or a reproducible task set.
- A route that returns an older model or an opaque fallback while carrying the new name.
- A one-day accepted-task win that disappears across repeated runs or the full observation window.
Write these rejection rules now, in the same document as the acceptance rubric. Teams that skip this step tend to redefine success around the first exciting result.
Evaluate Claude Fable 5.1 on EvoLink Track Gemini 4 API availabilityFAQ
Can anyone prove Gemini 4 costs less than Claude Fable 5.1 for agents today?
No. Google has not published a Gemini 4 price, cache rule, context window or API entry as of September 16, 2026, so any cost claim in either direction has no published denominator. Record your Fable 5.1 cost per accepted task now so a later claim has something to be checked against.
Can cache prices be compared across providers directly?
Not as single numbers. A cache read is comparable only when TTL, cache scope (prefix versus explicit), minimum cacheable size and Standard-versus-Batch mode are all matched. Anthropic documents Fable 5.1's 5-minute and 1-hour write tiers and its read rate; Gemini 4 has published none of these.
How do I count the cost of a failed long-running task?
Attach every attempt, resume and reviewer minute to the same task ID. Failed and unaccepted tasks stay in total spend but not in the accepted-task denominator. Report API spend per accepted task and review minutes per accepted task as two figures, never merged.
How should I record features Gemini 4 has not published?
Create the field with the value "Not published as of [date]". Do not fill it with a Gemini 3.x value, a leaked figure or zero. When Google publishes a contract, replace the placeholder and note the change date so historical comparisons stay honest.
How do I keep my verified Claude baseline while testing a new route?
Is Fable 5.1 available on EvoLink now?
claude-fable-5-1. Verify the returned model identity and usage counters on your first request before routing production traffic, and check current pricing on the Claude Fable 5.1 product page. Gateway availability is not a production verdict; that comes from your own replay, shadow and canary results.Sources
- Anthropic: Claude Fable 5.1 model overview
- Anthropic: Claude Fable 5.1 migration guide
- Anthropic: pricing, cache TTL and Batch
- Anthropic: Claude Fable 5.1 announcement
- Anthropic: Claude Fable product page
- Anthropic: model deprecations
- Google: Alphabet Q2 2026 earnings message from the CEO
- Google: Gemini 3.6 Flash announcement
- Google: Gemini API models catalog
- EvoLink: Gemini 4 API status
- EvoLink: Claude Fable 5.1
- EvoLink: Gemini 4 release tracker
- EvoLink: Claude Fable 5.1 vs GPT-6 Astra
