Accounting Cache Pricing: 4 Model Token Bills 2026

The 2026 state of frontier model cache pricing for accounting firms is a blended-bill problem, not a sticker problem. Claude Fable 5.1 and GPT-6 Astra both list at $10 input / $50 output per 1M tokens. That tie is real. It is also incomplete. Fable 5.1 cache reads are $0.25 per 1M. Astra cached input is $1.00 per 1M. Claude Fable 5 still caches at $1.00. GPT-5.6 Sol lists at $4 / $20 with cached input at $0.40. A firm that rereads the same binder, PBC folder, or payroll file all month is buying the cache column. A firm that pastes a new PDF every time is buying list input and should stop pretending otherwise.

This page is a census of those four public cards as of 3 Sep 2026, mapped onto accounting work. It is not a vs trophy. US Tech Automations is not a fifth model. It is mentioned only where a cache-warmed memo still has to wait on a person before it touches the ledger.

TL;DR

• Claude Fable 5.1 and GPT-6 Astra tie on list I/O at $10 / $50; they do not tie on cache: Fable cache reads: $0.25 / 1M versus Astra cache reads: $1.00 / 1M. A 5-minute Fable cache write is Fable 5m cache write: $12.50 / 1M.

• Claude Fable 5 is the control row: same $10 / $50 list as 5.1, cache still $1.00, which is why Anthropic can claim typical bills ~25% cheaper on 5.1 without cutting list I/O.

• GPT-5.6 Sol remains the cheap overflow at $4 / $20 and $0.40 cached input; AA Intelligence cost/task is $0.95 versus Astra $1.67 and Fable 5.1 $3.69.

• Blended cost follows hit rate. If cache never hits, Fable 5.1 and Astra cost the same to read new tokens. If the binder is warm, Fable 5.1 pulls away on the read column.

What the numbers say about cache pricing

according to Anthropic, $0.25 per million tokens is the Claude Fable 5.1 cache-hit price (0.025× base input), with 5-minute cache writes at $12.50 and 1-hour writes at $20.00, while Claude Fable 5 cache hits remain $1.00. according to OpenAI, $1.00 is GPT-6 Astra short-context cached input, with cache writes at $12.50, versus GPT-5.6 Sol cached input at $0.40 and writes at $5.00.

according to Microsoft Azure, $10 / $1 / $12.50 / $50 is Standard Global short-context Astra on Foundry, and long context is $20 / $2 / $25 / $75. US Data Zone is about 10% higher on that same post. according to BLS, 6 percent is the projected 2023–33 employment growth for accountants and auditors, which is why firms are buying draft labor from models instead of assuming a hiring wave will absorb PBC text.

according to Artificial Analysis, $1.67 is GPT-6 Astra Intelligence cost/task (max) versus $3.69 for Claude Fable 5.1 and $0.95 for GPT-5.6 Sol. Do not say Fable is cheaper than Astra on that column. Cache pricing and AA task dollars disagree, and a firm finance lead needs both.

Source line: Anthropic pricing page; OpenAI pricing page; Azure Foundry 3 Sep 2026 post; AA leaderboard 3 Sep 2026. Fast mode is not in this table: API docs 2× Standard, Help Center Codex/Work 2.5× Standard — name the surface on the invoice.

Why accounting operations break at cache scale

Accounting operations break at cache scale when the binder is huge, the question is small, and every run pays list input again. A PBC folder, a prior-year return, a chart of accounts, and a firm memo template are stable. The week's question is not. If you paste all four every time, you do not have a cache strategy. You have a paste habit.

Payroll is the clean example. The Gusto export shape does not change every Friday. The amounts do. See Gusto to Sage Intacct payroll journals. Invoice intake is the messy example: every PDF is new, so cache hits stay low and Sol's $4 input starts to look rational. See invoicing software cost for accounting firms. Outsourced production is the volume example: the same workpaper template should be cached once per engagement type. See workflow tools for outsourced accounting.

Astra long context above 272K input doubles input and cache and multiplies output 1.5× for the full request, except Codex skips that multiplier and does not bill cache writes. Dumping a 400-page binder into Astra without a cache plan is how a $10 sticker becomes a $20 sticker. Fable 5.1 prices the whole 1M window at standard per-token rates, which is a different long-file story, not a promise it is cheaper on AA tasks.

Fable 5.1 on AWS is a Covered Model with up to 30-day retention plus AWS human review unless you are EFS-eligible for ZDR through 31 Dec 2026. Cache savings do not override that line. Client files in a tax binder are exactly the content to ask about before you warm a 1-hour cache on Bedrock.

Staff still review. Cache makes the second run cheaper. It does not make the conclusion signed. A fluent memo that never ties to the trial balance is still a draft, whether it cost $0.25 or $1.00 to reread the binder.

The automation blueprint for a blended cache bill

Worked example

When a firm sets Anthropic cache_control on the engagement binder, Claude Fable 5.1 can hit that prefix at $0.25 per 1M while Claude Fable 5 and GPT-6 Astra still sit at $1.00 per 1M on their own cache columns, and GPT-5.6 Sol sits at $0.40. Official field: Anthropic prompt caching. Three figures that belong on the same recipe: Fable 5.1 cache hits $0.25, Fable 5.1 5-minute writes $12.50, Fable 5.1 1-hour writes $20.00. US Tech Automations is the workflow step that keeps cache_control on the binder, runs the memo job, and routes a reviewer hold before anyone posts a journal. It does not replace the four model cards.

A practical census for one CAS engagement looks like this. Write the chart of accounts and last year's variance template once (cache write). Ask this month's question many times (cache read). Attach only the new TB slice as uncached input. If you rebuild the system prompt every run, you pay write prices again and you invalidate Fable 5.1 thinking if you edit earlier turns. Treat the conversation as append-only. Moving cache_control markers is allowed; rewriting the prefix is how bills and thinking both break.

Astra callers that still send temperature will not be able to. Tools need the Responses API. Fast mode is a second price: 2× on API docs, 2.5× on Help Center Codex/Work. A "blended" invoice that mixes Fast mode, long context, and cache writes without naming the surface is not a census. It is a mess.

Sol remains the overflow when the file is new and short. You will not win a cache war against Fable 5.1's $0.25 if the prefix never repeats. You will win the uncached PDF pile at $4 input. That is the whole point of keeping four rows on the table.

Cost breakdown for a blended month

The table is arithmetic from the public cards, labeled as such. Assumptions: 20M input tokens, 80% cache hit, 4M output, short context, no Fast mode, no batch.

Source line: 20M / 80% hit / 4M output arithmetic from 3 Sep 2026 list prices. Cache writes omitted; a cold start that writes 16M at Fable 5.1's $12.50 5-minute rate is $200 before any read. Batch is half on the Anthropic and OpenAI cards when the memo can wait. Astra cached input: $1.00 / 1M is 4× Fable 5.1's read column at the same list I/O.

A second census for uncached invoice PDFs (10M input, 0% hit, 2M output): Fable 5.1 and Astra both $100 + $100 = $200; Fable 5 the same; Sol $40 + $40 = $80. Cache marketing does not apply. Sol wins that pile on dollars. Fable 5.1 wins the warm binder. Astra wins independent AA cost/task versus Fable ($1.67 vs $3.69) if you are buying scored intelligence, not rereads.

Source line: same public cards. At 0% hit, Fable 5.1 and Astra match. At 80% hit, Fable 5.1's $0.25 column is the only reason it undercuts Astra at identical list I/O. Sol stays cheapest at every hit rate in this mix because list I/O is 40% of the $10 / $50 pair.

US Data Zone Astra on Foundry is $11 / $1.10 / $13.75 / $55 short and $22 / $2.20 / $27.50 / $82.50 long. If the firm required a US data zone, the blended table above is the wrong sticker. Use the Foundry row.

Vendor / stack landscape for cache bills

Four models, two labs, one Azure invoice path. Chat seats are not in this census except as a warning: Plus and Claude Pro quotas are not $0.25 cache reads. API cache is.

Astra is not generally on ChatGPT on 3 Sep 2026. Enterprise Astra is off until an admin enables it. Fable 5.1 is live. Sol is callable. Fable 5 remains the control price for firms that have not moved the model id.

Pros and cons

Pros

• Claude Fable 5.1: cache reads $0.25; list I/O unchanged from Fable 5; Anthropic ~25% typical bill cut, up to ~45% on agent loops; 1M context at standard rates; live 1 Sep 2026.

• Claude Fable 5: known SKU; cache $1.00 matches Astra's read column, so a side-by-side with 5.1 isolates the cache cut; still $10 / $50.

• GPT-6 Astra: same $10 / $50 list; AA cost/task $1.67; Foundry Global short matches the OpenAI card; OSWorld minutes are a different job than cache.

• GPT-5.6 Sol: $4 / $20; cache $0.40; AA cost/task $0.95; promotional API pricing at least through 21 Nov 2026; best uncached PDF pile.

Cons

• Claude Fable 5.1: AA cost/task $3.69; 5-minute writes $12.50 and 1-hour writes $20.00; AWS Covered Model retention; forced tool_choice any/tool returns 400.

• Claude Fable 5: cache reads 4× 5.1; no $0.25 column; you are paying 5.1 list I/O without the cache cut if you have not migrated.

• GPT-6 Astra: cache $1.00; long context >272K doubles input/cache; Fast mode 2× or 2.5×; not generally in ChatGPT on 3 Sep 2026.

• GPT-5.6 Sol: lower independent knowledge-work ranking than Fable 5.1; slower OpenAI-run OSWorld minutes than Astra; not the cache-cut story.

FAQs

Do Claude Fable 5.1 and GPT-6 Astra cost the same?

They cost the same on list input and output: $10 / $50 per 1M. They do not cost the same on cache reads: $0.25 versus $1.00. They do not cost the same on AA Intelligence cost/task: $3.69 versus $1.67. A blended month depends on hit rate. A one-off paste ignores cache and still ignores the AA column.

Is Fable 5.1 cheaper than Fable 5?

On cache reads, yes: $0.25 versus $1.00, a 75% cut. On list I/O, no: both $10 / $50. Anthropic's ~25% typical-bill estimate (up to ~45% on agent loops) is a cache-mix claim, not a list-price cut. If your firm never hits cache, 5.1 is not cheaper.

Should an accounting firm default to GPT-5.6 Sol?

Default Sol when the file is new, short, and uncached, and when $0.95 AA cost/task is the ranking. Do not default Sol when the binder is stable and Fable 5.1's $0.25 reads would dominate. Do not default Sol when you needed Astra's computer-use minutes. Cache pricing is not a single winner.

How do Fast mode and long context change the census?

Fast mode: 2× Standard on API docs, 2.5× Standard on Help Center Codex/Work — write the surface on the bill. Long context for Astra: >272K input doubles input and cache and 1.5× output, except Codex. Sol's long-context card is $8 / $0.80 / $10 / $30. Either multiplier will swamp a $0.25 versus $1.00 argument if you dump the whole drive.

Does prompt cache replace a knowledge-management system?

No. Cache is a token discount on a prefix you already sent. It is not a workpaper repository, a PBC tracker, or a signed memo. Keep the binder in a real system. Use cache so the model can reread it without paying list input every question.

When is a workflow layer part of the cache bill?

When the cached binder, the memo, and a ledger draft must share a client ID and a reviewer. If partners only paste in Claude or ChatGPT, the cache column is still worth reading, but you do not need another product. Zapier, Make, or n8n can pass a file if you already own that recipe; this page does not claim they cannot retry or cannot log.

Key Takeaways

• Fable 5.1 and Astra tie at $10 / $50 list and split on cache: $0.25 versus $1.00.

• Fable 5 is the control row that proves 5.1's bill cut is the cache column, not list I/O.

• Sol is the uncached and AA-task-dollar cheap row at $4 / $20 and $0.95 per AA task.

• Blended cost is hit rate times those columns; 0% hit makes Fable 5.1 and Astra equal on reads.

• Fast mode and >272K Astra context are separate multipliers — name them or the census is fiction.

Who this is for

This page is for accounting firm COOs, CAS leads, and controllers who approve Claude or OpenAI invoices and keep seeing $10 / $50 treated as the whole story. Typical shape: a firm that rereads binders, payroll files, or workpaper templates and has not yet measured cache hit rate.

Red flags: skip this census if the firm does not use API tokens at all and only pays a chat seat; skip it if legal has already banned one lab; skip it if every file is a unique PDF and you already know cache will not hit; skip it if you wanted a two-product vs page — this is a four-row price census.

If the job is only "which chat tab," read the access pages and stop. If the job is "why is this invoice 4× last month's," pull hit rate, Fast mode surface, and long-context flags before you change model ids.

The workflow layer's homepage is US Tech Automations. Use it when a cached memo still has to wait on a person. See pricing only after the cache mix is measured, not guessed.