Claude Code vs OpenAI Codex: Ledger Scripts (2026)

Ledger automation is a coding-agent job, not a chat job. Claude Code and OpenAI Codex are the two agents an accounting firm can point at payroll-to-GL mappings, PBC request trackers, and close scripts. This page is only those two products. The question is which agent writes the script you will still review before it posts.

Claude Code is Anthropic's agent harness. OpenAI Codex is OpenAI's. Independent scoring on 3 September 2026 puts Claude Code at 70 on the Artificial Analysis Coding Agent Index and OpenAI Codex at 67. Codex is the cheaper token story on the new OpenAI flagship. Claude Code is the higher index. Neither agent should post a journal unsupervised.

TL;DR

• Pick Claude Code when the firm wants the higher independent Coding Agent Index score of 70 and will live with Fable-class token bills and thinking-block rules.

• Pick OpenAI Codex when the firm wants 67 on that same index, lower token use on the new OpenAI flagship (about one-third of the prior Codex flagship), and a path that already sits in ChatGPT Work/Codex for many seats.

• List I/O for the underlying frontier models is $10 / $50 per 1M tokens either way. Cache and Fast-mode math is not a tie.

• A script that posts to Sage Intacct or QuickBooks without a reviewer is a control failure, not a productivity win.

What the numbers say

Coding agents are scored as agents, not as chat models. according to Artificial Analysis, OpenAI Codex with GPT-6 Astra scores 67 on the Coding Agent Index, using about one-third the tokens of the prior Codex flagship. according to Artificial Analysis, Claude Code with Claude Fable 5.1 scores 70 on that index. Three points is a lead, not a wipeout. Token burn is the cost lever.

Claude Code Coding Agent Index: 70

OpenAI Codex Coding Agent Index: 67

The close calendar is why a three-point agent gap is not the whole decision. according to the Journal of Accountancy (2025), mid-market month-end still runs 8–10 business days. A script that saves 20 minutes on a payroll journal and then fails on the 9th day of a 10-day close is a late close, not a clever agent.

AA articles 2026-09-01 and 2026-09-03. List prices from Anthropic and OpenAI cards 2026-09-03. Help Center Fast mode is 2.5× on Codex/Work; API docs are 2×. Name the surface.

Astra is not generally on ChatGPT on 3 September 2026. Codex/Work is the OpenAI surface where a firm is more likely to meet the new flagship first. Claude Code on a paid Claude seat can run Fable 5.1 today. Do not plan a Monday close on a ChatGPT tile that is not there.

Why accounting operations break at scale

CAS and outsourced accounting break when the same journal is typed twice: once in Gusto or the payroll host, again in Sage Intacct or QuickBooks. Staff do that because the mapping lives in someone's head. Busy season deletes that headcount. according to AICPA in the 2025 PCPS CPA Firm Top Issues Survey, 62% of CPA firms report tech-workflow adoption. The remaining desks are not "waiting for agents." They are still re-keying.

The labor pool is large and still finite during 1040 weeks. according to BLS (2024), the United States has 1,579,800 accountant and auditor jobs. That count does not create spare close capacity in March. Scripts exist to move payroll, bills, and PBC requests without a senior touching every line. Agents exist to write those scripts faster. Controls exist so the scripts do not post fiction.

PBC tracking is the audit cousin of the same failure: requests live in email, evidence lives in a portal, and the binder is a zip file. If the firm is still chasing that in threads, fix the tracker before you pick an agent: audit PBC request tracking.

Scale is not "more CPAs." Scale is more entities, more payroll runs, more bank accounts, and the same 8–10 day close. An agent that rewrites a whole mapping file every run is a risk event. Anthropic's own Fable 5.1 docs warn that the model will rewrite whole files. Codex will happily keep going. The firm's standard is a diff, a reviewer, and a post in a sandbox.

The automation blueprint

The useful script is not "write me a payroll integration." The useful script is: payroll host finishes, webhook fires, mapping table applies, journal lands in a review queue, reviewer posts. Gusto to Sage Intacct is the canonical version of that path; see Gusto to Sage Intacct payroll journals.

Worked example: when Gusto emits payroll.processed, documented in Gusto's webhook events, US Tech Automations catches the event, asks Claude Code or OpenAI Codex to refresh the mapping script against 4 pay-group codes, and holds 1 journal with 18 lines in a review queue so a senior can check wages, taxes, and clearing before Sage Intacct posts. Three figures sit on that paragraph: 4 pay groups, 1 journal, 18 lines. The backticked token is the official event, not a homemade payroll.status.

Do not let the agent post. Let the agent draft the mapper and the test. The post is a human or a tightly scoped API call after the reviewer field is set. Outsourced shops that already run a workflow layer should keep the job board as the system of record; see workflow tools for outsourced accounting.

Claude Code will use Fable 5.1 thinking blocks. Editing earlier turns invalidates them. Forced tool_choice any/tool returns 400. Codex on Astra has no none reasoning and needs the Responses API for tools. Those constraints belong in the repo README, not in a partner Slack thread after a failed close.

Cost breakdown

Both underlying frontier lists are $10 / $50 per 1M tokens. Claude Code inherits Fable 5.1 cache reads at $0.25. Codex inherits Astra cache reads at $1.00. AA Intelligence cost per task is not a coding-agent invoice, but it is the honest independent burn ranking: Astra $1.67, Fable $3.69. Codex is the cheaper token story. Claude Code is the higher agent index.

Fast mode on OpenAI is 2× Standard in the API docs and 2.5× Standard on the Help Center Codex/Work card. If the firm turns Fast on for a close-week script, quote the surface. Batch is half of Standard on both labs.

Prices as of 2026-09-03 from Anthropic and OpenAI cards. AA task $ is Intelligence, not a Gusto journal quote. Do not say Fable is cheaper than Astra on the AA task column.

Astra AA Intelligence cost/task: $1.67

A close that still runs 8–10 days will not be rescued by a $0.25 cache read. It will be rescued by a mapper that does not require a senior to re-key 18 lines. Tokens are the small bill. Reviewer hours are the large bill.

Vendor / stack landscape

Claude Code and OpenAI Codex are agents. They are not the GL. They are not Gusto. They are not the firm's workflow board. The stack is payroll host + GL + review queue + agent. Pick the agent the way you pick a senior: who writes a safer first draft, and what do they cost.

1 = yes / present; 0 = no. Index and $ from AA and list cards 2026-09-03.

How we evaluated

Weights assume a CAS or outsourced accounting desk that will review every post. A product shop writing internal glue can raise token cost and lower control. A PCAOB-adjacent audit shop should raise hold-and-log and treat the index as a draft-quality hint.

Pros and cons

Claude Code

Pros

• Independent Coding Agent Index 70, the higher of this pair.

• Fable 5.1 cache reads at $0.25 per 1M, which helps a looping mapper that rereads the same chart of accounts.

• Paid Claude seats can run it on 3 September 2026 without waiting on an OpenAI ChatGPT enable.

• Stronger published showing on long knowledge-work benches for the underlying model, which helps when the script has to read a 40-page PBC list.

Cons

• Underlying AA Intelligence cost per task is $3.69, more than Codex's underlying $1.67.

• Fable 5.1 thinking blocks break if you edit earlier turns; earlier models cannot read them.

• Anthropic warns Fable 5.1 rewrites whole files, which is a mapping-table hazard.

• Forced tool_choice any/tool returns 400, so naive tool glue will fail closed.

OpenAI Codex

Pros

• Independent Coding Agent Index 67, close enough that token burn can decide the bake-off.

• Underlying Astra AA task $ is $1.67, and Codex token use on the new flagship is about one-third of the prior Codex flagship.

• Work/Codex is the OpenAI surface where firms will meet Astra first while ChatGPT remains gated.

• Codex does not add Astra's long-context multiplier above 272K and does not bill cache writes, which matters if a close file is huge.

Cons

• Cache reads on the underlying model are $1.00 versus $0.25.

• Fast mode is 2× on API docs and 2.5× on the Help Center Codex/Work card; a close-week toggle can erase the token win.

• Astra is not generally on ChatGPT on 3 September 2026, so "we'll just use ChatGPT" is the wrong access story.

• No none reasoning and Responses API for tools, same class of build friction as Fable's thinking rules.

FAQs

Which agent should write our Gusto-to-Intacct mapper?

Claude Code if you want the 70 index and can afford Fable-class tokens. OpenAI Codex if you want 67 and the cheaper underlying task $. Both still need a reviewer before Sage Intacct posts.

Is OpenAI Codex cheaper than Claude Code?

On the underlying AA Intelligence task column, the Astra-backed agent is $1.67 versus $3.69. On cache-heavy loops, Claude Code's $0.25 cache read can invert a bill. List I/O is $10 / $50 both ways.

Can we let the agent post journals after a clean test?

No. A clean test is one payroll. The 18-line journal still needs a senior on wages, taxes, and clearing. Agents draft. Reviewers post.

Does Zapier replace either agent?

No. Zapier, Make, or n8n can copy a finished payroll into a folder or ping Slack. That is useful when copy is the whole job. It is not a mapper, a diff, or a review queue. Use those tools for the copy. Use an agent for the script. Use a workflow layer when the event, the script, and the hold must be one run.

What if Astra is not in our ChatGPT workspace yet?

Use Codex/Work if that seat has the new flagship, or use Claude Code on a paid Claude seat. Do not stall a close on a ChatGPT tile. Enterprise Astra stays off until an admin enables it.

Is METR's time horizon the tie-breaker?

No. METR 50%/80% horizons are not published for either underlying model as of 3 September 2026. Use the Coding Agent Index, your own review-hour log, and whether the script has a hold.

Key Takeaways

• Claude Code 70 versus OpenAI Codex 67 on the independent Coding Agent Index is a real lead, not a rout.

• Underlying list is $10 / $50 both ways. Cache $0.25 versus $1.00 and AA task $3.69 versus $1.67 are the cost split.

• Fast mode on Codex/Work is 2.5× on the Help Center and 2× on API docs. Name the surface.

• payroll.processed to a review queue is the blueprint. Agents do not post.

• US Tech Automations belongs when the webhook, the agent, and the hold are one workflow. Zapier, Make, or n8n is enough when the job is a copy.

Who this is for

This page is for CAS partners, outsourced accounting ops, and firm IT that already run Gusto or another payroll host plus QuickBooks or Sage Intacct, and that want an agent to write the mapper. It is not for a partner who pastes one CSV a month. It is not for a cyber waitlist.

Red flags: no sandbox GL; no reviewer field; letting the agent commit to main; waiting for a ChatGPT tile to start the mapper; treating Zapier as the GL.

When NOT to use US Tech Automations: if the firm has 10 payroll journals a year and a senior already keys them, keep the senior. If Zapier, Make, or n8n already copies payroll.processed into Slack and a human posts, keep that copy. If Intacct or QBO native rules already create the journal from a CSV the payroll host drops, do not add an agent. Add orchestration when the event must refresh a script, open a review queue, and only then post.

Pick Claude Code for the higher agent index and the cheaper cache. Pick OpenAI Codex for the cheaper underlying task $ and the Codex/Work path. Route the hold through US Tech Automations only when payroll, script, and reviewer have to be one run; the shape is on agentic workflows.