Copper and blue glass workflows with memory tiles and a recovery loop for Gemini 4 Argon vs Claude Fable 5.1

analysis

Gemini 4 Argon vs Claude Fable 5.1: Agent Cost & Quality

Q: Is Fable 5.1 available on EvoLink now?

EvoLink has a configured route named 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.

EvoLink Team

Product Team

September 16, 2026

Updated on October 13, 2026

17 min read

Compare Gemini 4 Argon and Claude Fable 5.1 by accepted long-running tasks, including cache use, retries and human review. Argon is available through EvoLink for evaluation with your existing key. Keep Fable 5.1 serving the workflows it already handles well while you measure whether Argon improves the complete task.

Your workflow	Starting choice	What would justify a change
Long reports or code deliverables need more output headroom	Evaluate Argon with an explicit output budget	More accepted complete deliverables, without excessive latency or review
Your Fable agents rely on its thinking blocks and cache behavior	Keep Fable as the baseline	Argon preserves task completion and reduces combined operating burden
Repeated prefixes dominate token usage	Test both with your actual prefix reuse pattern	Observed cache hits and total task spend, not a nominal discount
Tasks often need retries or human rescue	Compare full traces on both routes	Fewer failed loops, recoveries and reviewer minutes

Use the Argon model page and Fable 5.1 model page for current access and prices. This guide focuses on the long-agent decision; see the release timeline, Astra comparison or Fable versus Astra for those separate questions.

Decision flow from model announcement through API documentation and route verification to workload evaluation

Status and Evidence Table

Question	Gemini 4	Claude Fable 5.1	Safe decision
Official status	Announced · Restricted rollout	Claude Fable 5.1 · 2026-09-01	Nominal token rates and provider evaluations can be compared now; a matched EvoLink workload test remains pending.
Provider API model ID	Not established by the announcement; use the configured EvoLink ID on its model page for gateway requests	`claude-fable-5-1` (Anthropic)	Use the selected route’s documented ID; keep provider and gateway identifiers separate
Context and max output	Google announces an output limit of 1M tokens for extended reasoning and generation. This is an output limit, not a confirmed input context window or an EvoLink parameter guarantee.	1M context; 128K max output (Anthropic)	Record how much of the window your traces actually consume
Input modalities	Supported public API input modalities still await documentation.	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	Google's announced introductory rates are $2 per 1M input tokens and $10 per 1M output tokens. After the introductory period they become $4 and $20 respectively. The end date is not announced. These are Google rates, not an EvoLink quote. Google announces a 95% discount on cached input tokens: $0.10 per 1M at the introductory input rate, calculated as $2 × 5%.	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	Available through EvoLink; model ID and current pricing on the Gemini 4 page	Configured route `claude-fable-5-1`; pricing on the product page	Verify returned model identity on the first request of any route

Google's published evaluations provide an initial comparison, not a matched EvoLink test. They do not establish latency, accepted-task cost or a universal winner. The table below includes both a favorable result and a result where Argon trails the comparison model.

Google-reported evaluation results

Benchmark	Gemini 4 Argon	Claude Fable 5.1
DeepSWE v1.1	77.9%	67.4%
FrontierSWE v2	55.0%	56.3%

Source: Google's model evaluations and methodology. Provider-reported scores; settings and task definitions matter. No EvoLink head-to-head run is claimed.

Read the settings before interpreting the gap. Google's methodology uses pass@1 unless noted and generally the highest available thinking settings. DeepSWE combines an Argon run using mini-swe with comparison results from public leaderboards or system cards; FrontierSWE comes from the Proximal leaderboard. These are not two models rerun inside one EvoLink harness. Keep benchmark version, harness, reasoning setting and source beside each score, and do not translate percentage-point gaps into a predicted production success rate.

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 Argon per class as well.

Workload	What "accepted" means	What to record on Fable 5.1 now	What the Argon evaluation must 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 fixes what “the same task” means before you compare two routes.

Context Reuse and Cache Conditions

Long-running agents repeatedly send repository context, tool schemas and instructions. The announced cache discount supports an initial price calculation, but a full cache-cost comparison still needs matching operating conditions.

Google announces a 95% discount on cached input tokens: $0.10 per 1M at the introductory input rate, calculated as $2 × 5%. Cache write/storage charges, TTL, eligibility, post-introductory cache terms and gateway support still need a documented API contract. A five-minute cache and a one-hour cache are not interchangeable, nor are automatic prefix caches and explicit shared cache objects. Compare the same TTL, scope, minimum size and service mode before inferring savings.

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 changes can prevent reuse of the affected cached prefix;
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.

Keep those fields per trace, not per month. For Argon, measure cached usage under the selected EvoLink route and check the documented cache behavior; do not assume Anthropic’s TTL or write semantics transfer to Gemini. EvoLink rates for Fable 5.1 are on the product page.

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. Malformed calls can indicate a model or integration problem; distinguish their causes from provider and tool failures before attributing a difference to the model.
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	Google's announced introductory rates are $2 per 1M input tokens and $10 per 1M output tokens. After the introductory period they become $4 and $20 respectively. The end date is not announced. These are Google rates, not an EvoLink quote.
Cache write	Tokens written, split by 5-minute and 1-hour TTL, with the write count per task	Record from the selected route documentation and actual task traces
Cache read	Tokens served from cache per task; also log hit ratio (cache read / (cache read + fresh input))	Google announces a 95% discount on cached input tokens: $0.10 per 1M at the introductory input rate, calculated as $2 × 5%. Cache write/storage charges, TTL, eligibility, post-introductory cache terms and gateway support still need a documented API contract.
Output tokens	Actual billed output usage across every call, including reasoning where billed; not just visible answers and tool arguments	Google announces an output limit of 1M tokens for extended reasoning and generation. This is an output limit, not a confirmed input context window or an EvoLink parameter guarantee.
Tool calls and external cost	Count and any metered cost (search, code execution, third-party APIs) per task	Record from the selected route documentation and actual task traces
Retries	Attempts beyond the first, with tokens per attempt, attached to the same task ID	Record from the selected route documentation and actual task traces
Fallback route usage	Whether any turn was served by a different model, and which one, verified by the returned `model` field	Record from the selected route documentation and actual task traces
Human review minutes	Reviewer and operator time per task, recorded as minutes	Measure on your own task runs; not a vendor-published figure
Wall-clock to acceptance	Time from task start to accepted result, including waits and human turns	Measure on your own task runs; not a vendor-published figure
Accepted-task rate	Accepted tasks / attempted tasks per workload class over the window	Measure on your own task runs; not a vendor-published figure

Cost per accepted task is API spend divided by accepted tasks, reported next to review minutes per accepted task. Lower API spend alone does not establish lower total cost. Include review time at your own labor rate before deciding whether the route saves money.

Announced unit prices support budget scenarios, not a savings guarantee. For an illustrative 1M uncached input tokens plus 100K output tokens, Google's Argon rates imply $3 during the introductory period and $6 afterward. At $10/$50 per 1M input/output, the comparison model's standard base cost is $15. This arithmetic excludes caching, tools, long-context surcharges, retries, taxes and gateway differences, and assumes equal token counts rather than equal completed work. Actual workload tokenization and reasoning use can differ. Budget both Argon periods because the introductory end date is unknown; verify EvoLink rates separately. This example aggregates tokens across multiple requests. Argon long-context billing rules still need confirmation.

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.

Run a workload-specific comparison through EvoLink while preserving your current baseline, acceptance criteria and fallback during replay, shadow and canary stages. API access enables the evaluation; it does not establish a quality or cost win.

Shadow Traffic and Rollback Plan

Introduce Argon 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.

"Returned model identity" verification means reading the model field and usage counters in every response and checking they match the configured route. On a gateway, this checks consistency with the configured route; the returned label alone does not independently prove the underlying model identity. Reconcile it with route records and billing, and investigate unexplained fallbacks. 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 Evaluate Gemini 4 Argon on EvoLink

FAQ

Can anyone prove Gemini 4 costs less than Claude Fable 5.1 for agents today?

Not for your production workload yet. Announced rates are lower than the comparison model's standard base rates, but route access, actual tokens, cache eligibility, retries and accepted-task quality must be measured before claiming total savings.

Can cache prices be compared across providers directly?

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 separately. Combine them into a total only after specifying the labor rate used.

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?

EvoLink has a configured route named 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.

Can I reuse a Fable 5 integration unchanged for Fable 5.1?

Do not infer compatibility from the shared family name. Verify the exact model ID, current thinking and tool-response requirements, and replay stored conversation turns against the Fable 5.1 documentation. Its adaptive thinking behavior and thinking-block handling need version-specific checks; an old transcript or parameter set is not automatically a valid migration fixture.

Sources

Updated October 13, 2026. Provider announcements and methodology were rechecked; provider-native features must be checked against the selected gateway route. Google-reported evaluations are not EvoLink tests.

All Posts

#Gemini 4 #Gemini 4 Argon #Claude Fable 5.1 #Google #Anthropic #Model Comparison