Fable 5.1 vs Opus 5: Coding Agent Score of 70 (2026)

A coding agent for a SaaS company is not a chat window that pastes a diff. It is the loop that reads a failing check, edits the repo, opens or updates a pull request, and stops for a human when the change touches billing, auth, or a public API. Claude Fable 5.1 and Claude Opus 5 are the two Anthropic models that actually show up in that loop in September 2026. This page compares those two products on the coding-agent harness, not on generic “which model is smarter” slides.

Claude Fable 5.1 vs Claude Opus 5 is a comparison of two Claude models used as the worker inside Claude Code (and similar repo agents), judged on independent coding-agent scores, list token prices, cache behavior, and the cost of keeping a merge behind a named reviewer. Neither model is your Git host. Neither is a substitute for a written merge policy.

TL;DR

• Choose Claude Fable 5.1 when the job is long-running agentic coding and you will pay Fable list rates ($10 input / $50 output per million tokens) to chase the higher Coding Agent Index.

• Choose Claude Opus 5 when the same Claude Code harness is enough and you want Opus list rates ($5 input / $25 output) on the majority of pull requests.

• Keep the harness honest: log stop_reason, token usage, and the GitHub pull_request id on every run. Do not treat a green check as a billed, reviewed change.

• US Tech Automations belongs only when a coding-agent run must cross GitHub, billing flags, and a human hold. Native Claude Code is enough when one repo and one reviewer already are the process.

What the numbers say

Independent coding-agent scores and Anthropic list prices are the two numbers that should sit on the same sheet. Mixing a provider demo with a seat quote is how teams pick the louder model and then overpay on quiet refactors.

Fable 5.1 CAI score: 70 according to Artificial Analysis, 70 on the Coding Agent Index when Fable 5.1 is run in Claude Code (counted 2026-09-01 / 2026-09-03). That is the independent coding-agent figure this page uses. It is not a reason to skip a reviewer.

Opus 5 list price: $5 / $25 according to Anthropic Claude pricing, $5 per million input tokens and $25 per million output tokens for Claude Opus 5, with cache hits at $0.50 per million. Fable 5.1 stays at $10 / $50 list with cache reads at $0.25. Same family, different meter.

Software hiring pressure is the operating backdrop, not a model score. Employment of software developers is projected to grow 17 percent from 2023 to 2033 according to BLS Occupational Outlook Handbook, 17% over that decade. A coding agent that saves review time is still competing with a market that is adding developer seats.

SaaS unit economics still set the ceiling for how much agent spend you can hide in COGS. Median SaaS net revenue retention in the $10–50M ARR band is 110% according to Bessemer’s State of the Cloud (2024), 110% median NRR in that mid-market band. Agent token bills that quietly raise COGS without raising NRR fail that test.

Source note: CAI 70 is Artificial Analysis for Fable 5.1 in Claude Code. Opus 5 is the cheaper list SKU on Anthropic’s public pricing table; do not invent a separate CAI cell for Opus 5. AA’s Astra-in-Codex 67 is a different harness and is not a third product on this page.

Median ARR per FTE in the $5–20M SaaS band is $145K according to ChartMogul (2024), $145K. Use it to size how many coding-agent dollars you can spend per engineer before the agent is a headcount substitute in disguise.

How we evaluated coding-agent harnesses

Weights assume a SaaS platform team that already ships through GitHub pull requests, has a CI check, and will not merge billing or auth changes without a human. A solo prototype repo should raise “list price” and lower “reviewer hold.”

A harness is the recorded path from event to patch to merge, not the model card. Fable 5.1 can be the better coder on CAI 70 and still be the wrong SKU if every small fix is billed at $10 / $50 and nobody logs cache reads. Opus 5 can be the right default worker if the same Claude Code recipes already pass your tests and you only escalate to Fable when the eval says the cheaper model stalled.

Why SaaS engineering operations break at scale

SaaS engineering does not break because someone picked the “wrong” logo. It breaks when three systems disagree about the same change: the git host thinks the pull request is mergeable, the agent thinks the task is done, and billing still exposes a flag the patch was supposed to gate. Feature-flag approvals, license-renewal reviews, and helpdesk macros all sit next to that seam; see feature-flag access approvals, helpdesk software for SaaS, and product analytics tools for SaaS for the non-git half of the same company.

The failure mode at 20 engineers is a coding agent that opens 40 pull requests a week, half of them drive-by refactors, with no unique key tying the agent run to the GitHub number. The failure mode at 80 engineers is the same agent with permission to touch billing and auth because someone copied a repo token into a shared secret. Neither Fable 5.1 nor Opus 5 will refuse that architecture for you. The model will keep writing patches until the bill or the incident does.

Long-running agents also change the cost shape. Fable 5.1 is priced for demanding agentic coding: thinking is always on, cache reads dropped to $0.25 per million, and Anthropic’s own docs tell you to start with Opus 5 for most workloads and move to Fable when Opus at higher effort still falls short. That sentence is the operating rule. Teams that default every repo to Fable 5.1 recreate the “we bought the flagship and now the quiet work is a luxury good” problem.

Protected paths need a second control. A coding agent that can edit pricing.ts, webhook handlers, or anything that changes a customer’s billable quantity is not a style-nits bot. Put those paths on a CODEOWNERS hold. Log the model id that produced the patch. If you cannot answer “did Fable 5.1 or Opus 5 write this file?” six weeks later, you do not have a harness. You have a chat log.

The automation blueprint

The blueprint is event → model → patch → review → merge, with a stop that a human owns. GitHub remains the system of record for the change. The model is the worker. Claude Code (or an equivalent repo agent) is the runtime. Your job is to decide which of the two Claude SKUs is the default worker and which events escalate.

Step 1: name the default model per repo. Most product and internal tools repos should default to Claude Opus 5. Escalate to Claude Fable 5.1 when an eval on the same tests fails twice or when the task is a multi-hour agent loop (large migrations, multi-package refactors). Write that rule in AGENTS.md. Do not leave it as Slack folklore.

Step 2: subscribe to GitHub pull_request events and require a unique run id. The agent may update the same PR many times. Each update is a new usage row. If you cannot join usage to pull_request.number, you cannot explain the month’s token bill.

Step 3: block merge on named paths until a reviewer with CODEOWNERS rights approves. The agent may open the PR. The agent may not satisfy the hold. That is the difference between a coding assistant and an unattended merge bot.

Step 4: export usage nightly. Fable 5.1 cache reads at $0.25 only help if the prefix is stable. Teams that rebuild the system prompt every turn pay cache writes and then wonder why the 75% cache cut never showed up.

Step 5: keep a fallback path that is not “retry Fable forever.” If Fable 5.1 returns stop_reason: "refusal", record it and decide whether Opus 5 should retry. Do not silently swap models mid-conversation without logging the swap; Fable 5.1 thinking blocks are not readable by earlier models.

Worked example

An illustrative SaaS platform team runs 28 backend engineers, 1,200 pull requests per month, and a 110% NRR target. When GitHub delivers a pull_request webhook with "action": "synchronize" (documented on GitHub’s pull_request webhook payload), a configurable US Tech Automations workflow can require three facts before the coding agent is allowed a second turn: a numeric pull_request.number, a head.sha of 40 hex characters, and a path list that does not include /billing or /auth. If those match, the run may call Claude Opus 5 at $5 / $25 per million tokens; if the same tests fail twice, the run may escalate to Claude Fable 5.1 at $10 / $50. Prerequisites: a GitHub App with pull_requests:write limited to non-protected paths, an Anthropic usage export, and a reviewer group on CODEOWNERS. Outputs: a run id, a model id, token totals, and a pass/fail reason — not a promised CAI lift. Three concrete figures in this path: 1,200 PRs per month, $5 versus $10 input, and CAI 70 as the Fable 5.1 ceiling rather than a guarantee on your repo.

Cost breakdown

List prices are not the bill. The bill is input + cache writes + cache reads + output, times how verbose the agent is, times how often it retries. Opus 5 is half of Fable 5.1 on list input and output. Fable 5.1 is cheaper on cache reads ($0.25 vs Opus 5’s $0.50). The crossover is a function of how much of the prefix actually hits cache.

Source note: unit prices from Anthropic’s public pricing table (checked 2026-09-03). The 10M / 2M / 6M volumes are an illustrative mix for a mid-size agent repo, not a measured customer bill. Cache writes ($12.50 per million on both Fable 5.1 and the Fable 5 family; Opus 5 writes at $6.25 per million for 5-minute cache) sit on top of this sheet when the prefix is unstable.

A team that puts Fable 5.1 on every nits PR will pay the $10 / $50 meter for work Opus 5 was designed to take. A team that never escalates will leave CAI 70 on the table for the migrations that actually need a long-horizon agent. The cost owner’s job is the mix, not the logo.

Batch is half of Standard on Anthropic’s table. That helps overnight evals. It does not help a developer waiting on a PR. Do not quote batch prices as the interactive harness rate.

Vendor / stack landscape

Only two products sit on this vs page: Claude Fable 5.1 and Claude Opus 5. Claude Code is the runtime both use. GitHub is the event source. US Tech Automations is an orchestration layer above that pair when a human hold and a unique run id have to live outside the model vendor.

Source note: prices and the “start with Opus 5” guidance are Anthropic docs (1 Sep 2026 Fable 5.1 pack). CAI 70 is Artificial Analysis. The “1 / 0” cells are presence flags from those docs, not quality scores.

Do not add GPT-6 Astra or Codex as a third column here. Astra’s Coding Agent Index of 67 in Codex is a useful external benchmark for “is 70 a lot,” and it is a different product, a different harness, and a different vs page.

Pros and cons

Pros

• Claude Fable 5.1: Independent CAI 70 in Claude Code; $0.25 cache reads; 1M context; built for long-running agentic coding when Opus 5 at high effort still fails the same tests.

• Claude Opus 5: $5 / $25 list is half of Fable on input and output; Anthropic’s own “start here” SKU for most workloads; same Claude Code recipes; cheaper default worker for nits and well-specified tickets.

Cons

• Claude Fable 5.1: $10 / $50 list; thinking always on so quiet tasks still burn reasoning tokens; forced tool_choice any/tool returns 400; earlier models cannot read its thinking blocks, so a sloppy fallback drops reasoning.

• Claude Opus 5: No independent CAI 70 to quote; cache reads at $0.50 are four times Fable 5.1’s $0.25, so hot agent loops can erase some of the list-price win; teams that never escalate will stall on long-horizon migrations.

FAQs

Should a SaaS team default Claude Code to Fable 5.1 or Opus 5?

Default to Claude Opus 5. Escalate to Claude Fable 5.1 when the same tests fail twice or the task is a long-horizon agent loop. Anthropic’s Fable 5.1 docs say to start with Opus 5 for most workloads.

Is CAI 70 a reason to put Fable 5.1 on every repository?

No. CAI 70 is a Claude Code eval, not your merge rate. Use it to justify an escalation path, not a default SKU.

How should we log a coding-agent run?

Store the GitHub pull_request.number, head.sha, model id, token totals, and stop_reason. If you cannot join those five fields, you cannot explain the bill or the patch.

Do cache reads make Fable 5.1 cheaper than Opus 5 overall?

Not on list input and output. Fable 5.1 cache reads are $0.25 versus Opus 5 at $0.50, but Fable input and output are still $10 / $50 versus $5 / $25. Run your own mix before you claim a winner.

When NOT to use US Tech Automations?

Skip it when Claude Code plus CODEOWNERS already is the process, when one repo and one reviewer cover every merge, or when a Zapier, Make, or n8n scenario already notifies engineering on pull_request and you trust that run log.

What breaks if we swap Fable 5.1 to Opus 5 mid-thread?

Fable 5.1 thinking blocks are not readable by earlier models. The API drops them. Log the swap and treat it as a new run rather than a silent continuation.

Key Takeaways

• Claude Fable 5.1 is the long-horizon coding-agent SKU (CAI 70 in Claude Code, $10 / $50 list, $0.25 cache reads).

• Claude Opus 5 is the default worker ($5 / $25 list) and the SKU Anthropic tells you to start with.

• A harness is pull_request id + model id + tokens + reviewer hold, not a model announcement.

• Do not compare these two against Astra-in-Codex on this page; 67 CAI is a different product and a different harness.

• Orchestrate across GitHub and billing paths only after unique run ids and a named reviewer exist.

Who this is for

This comparison is for a SaaS engineering manager or platform lead choosing the default Claude model inside a coding-agent harness, with a named owner for merge policy and token bills. It assumes you already host git on GitHub (or an equivalent that emits pull-request events) and you will not let an agent merge billing or auth paths unattended.

Red flags: skip a custom orchestration layer when Claude Code plus CODEOWNERS already runs the only required path, when you have no usage export, or when nobody will own protected-path reviews. Do not buy Fable 5.1 as a personality upgrade for nits. Do not keep Opus 5 as the only SKU if your evals on hard migrations already fail twice.

Zapier, Make, or n8n can take a pull_request webhook, retry a failed post to Slack, and keep a run log if you design observability, idempotency, access, and retention. That is a fair DIY choice for one stable notify recipe. A proposed agent design would add a durable run-id ledger and a human hold before merge on protected paths — not a claim that no-code cannot retry.

When NOT to use US Tech Automations: leave it out when the Claude Code default plus GitHub branch protection already is the process, when an iPaaS recipe already pages the reviewer with the PR url, or when the team will not staff a usage export. Honest self-selection beats a second platform fee.

The team at US Tech Automations can map a configurable pull_request → model → reviewer trail. Review agentic workflow pricing after you have named the default Claude SKU, the protected paths, and the reviewer group.