Engineers compare Fable 5, Codex, and local models on token bills

• Anthropic's launch post puts Claude Fable 5 at $10 per million input tokens and $50 per million output tokens, double the list price Simon Willison quoted for the Opus 4.x line in his first-impressions post.
• According to fresh HN discussion, engineers are describing Fable 5 as better at high-level reasoning, plan review, and ugly debugging sessions than Opus 4.8, while the main HN discussion roundup keeps circling back to the same catch, it burns through budget fast.
• In Simon Willison's product-market-fit post, Willison estimated roughly $2,180.16 in 30-day Claude Code and Codex usage from a $200 combined Max and Pro subscription spend, which is why flat monthly plans still look like a bargain for heavy agent users.
• The local-model camp in the HN summary of Local AI Needs to be the Norm is making a different argument entirely: route narrow tasks to smaller local models, keep tool use and retrieval close to the machine, and escalate only the hard calls to frontier APIs.
• Anthropic buried two expensive footnotes outside the headline pricing, a new classifier-fallback path and a mandatory 30-day retention policy for Mythos-class traffic, in its model docs and launch post.

You can read Anthropic's launch post, skim the API model guide, and then compare that vendor framing with Simon Willison's hands-on writeup, his earlier coding-agent cost post, and the HN local-first argument. The weird bit is that all three camps are talking about the same thing now: not benchmark supremacy, but how many expensive agentic runs you can afford before the model choice stops penciling out.

Anthropic shipped Claude Fable 5 and Claude Mythos 5 as the same underlying model with different access and safeguards. Fable 5 is public through the API, while Mythos 5 starts behind Project Glasswing and a trusted-access program.

The public sticker price is simple:

• Input: $10 per million tokens
• Output: $50 per million tokens
• Context window: 1 million tokens
• Max output: 128,000 tokens

Anthropic's model guide confirms the same numbers. In Simon Willison's first-impressions post, Willison notes that this is twice the Opus 4.5 through 4.8 price band, even before you get to any long-running agent loop.

The strongest pro-Fable argument in the evidence pool is not a benchmark chart. It is engineers saying the model changes how they work through a hard problem.

Three recurring claims show up in the HN thread and Willison's post:

• Better high-level reasoning and plan review than Opus 4.8, per fresh HN discussion
• More willingness to pivot toward larger simplifications instead of local patching, according to the HN discussion roundup
• Stronger long-horizon debugging, including the report in fresh HN discussion that it built a whole test harness around a Windows lifecycle bug before proposing fixes

That is Christmas-come-early material for coding-agent nerds. It also sounds expensive on purpose: the model is being praised for doing more work, over longer traces, with bigger context and more elaborate self-checking.

One reason sticker price is not the whole story is that many heavy users are not paying list API rates for every run. They are getting frontier-model usage through monthly plans, then back-solving what that usage would have cost at token rates.

In Simon Willison's post, he ran ccusage on his laptop and estimated about $2,180.16 in Claude Code and Codex tokens over 30 days while paying $100 for Anthropic Max and $100 for OpenAI Pro. The HN thread around that post adds the same pattern from other users: these tools chew through credits on test suites, compiler runs, and large agentic loops, but the perceived value can still beat pay-as-you-go if you live in them all day.

That is why the main Fable 5 HN thread reads less like a launch thread and more like a cost-accounting session. Engineers are comparing models by completed work per dollar, not by who won one more vendor benchmark.

The local-first counterargument is not that frontier models are weak. It is that too many product teams are sending cheap tasks to expensive models because chat APIs are the default abstraction.

The local-first stack described in Local AI Needs to be the Norm and its HN thread breaks that into a routing problem:

• Use small local models for narrow tasks that only need to be good enough
• Lean on tool use and retrieval to reduce hallucinations instead of scaling the base model first
• Keep privacy-sensitive work on device when possible
• Route only the hard, frontier-intelligence calls to cloud models

HN commenters add two practical constraints to that story. According to the HN discussion of local-first design, standardized OS or browser model APIs would make hybrid routing easier, while the same thread's hardware discussion points out that memory, VRAM, and latency still make serious local inference painful for many teams.

Anthropic's own docs surface two less glamorous details that showed up in the HN criticism almost immediately.

First, Fable 5 can refuse requests through new safety classifiers, and Anthropic's cookbook guide says integrators need to handle direct refusals, optional server-side fallback, and revised billing behavior. Anthropic also says blocked requests are not billed for input tokens, and fallback requests can be billed like cache reads in some flows.

Second, Anthropic's launch post says Mythos-class models now require 30-day retention for all traffic on first- and third-party surfaces. Fresh HN discussion flagged that policy friction alongside benchmark skepticism on day one, which means the bill engineers are comparing is not only dollars and tokens. It also includes which requests can be sent at all, and under what data-handling terms.