Freelance translator AI stack 2026: where post-editing earns and where it cannibalises your rate

The MTPE rate-cannibalisation question is simpler than most translator forums make it. You are asked to post-edit machine translation at 40–60% of your usual per-word rate. The only question worth answering before accepting is whether your word output in post-editing mode is high enough to hold your hourly income. If not, the work is subsidising the agency’s MT investment at your expense.

The maths require one number you likely do not track: your actual post-editing word output per hour, measured against your conventional translation output per hour in the same language pair and domain. Without that ratio, every MTPE rate discussion is an intuition contest you will usually lose.

Rate-card baseline: what the market pays by pair

Current full-translation and MTPE rates by major language pair, drawn from Slator’s 2024 Language Industry Intelligence report and cross-checked against ATA’s 2023 Compensation Survey for US-market comparators and ITI rate guidance for UK practitioners.

Sources: Slator Language Industry Intelligence 2024; ATA Compensation Survey 2023; ITI rates guidance. Regional variation is significant; rates above are indicative for agency work. Direct-client rates typically run 20–40% above the agency figures.

CSA Research’s 2023 MTPE adoption report found that 87% of language service providers now offer MTPE as a standard workflow option, up from 67% in 2020. The same data showed that post-editing rates have compressed 8–12% in nominal terms since 2021 as MT output quality improved. That compression is ongoing.

The EN → ES cell in the table is the most contested. Spanish is MT’s strongest major language pair: DeepL, Google Translate, and the major LLMs all produce commercially usable Spanish from English at a significantly higher baseline than they produce Japanese or German. That quality lift is why EN → ES MTPE rates are under the most downward pressure. If EN → ES is your primary pair and you are below 0.08 €/word full rate, the MTPE rate band (0.03–0.07 €/word) puts you in territory where the break-even math is especially unforgiving.

The 1.8× productivity threshold: the break-even worked example

Take a translator working EN → DE at 0.12 €/word, producing 400 words/hour on conventional translation. Baseline hourly income: 48 €/hour.

An agency offers MTPE at 0.06 €/word: 50% of the full rate.

To hold 48 €/hour at 0.06 €/word, the translator needs to post-edit at 800 words/hour. That is a 2.0× throughput requirement against a 400 words/hour conventional baseline.

The 1.8× figure in the primary claim is the practical threshold, not the theoretical break-even. The difference accounts for the cognitive overhead of switching between post-editing mode and translation mode in a single working day. Post-editing is not simply faster translation: it requires suppressing your own phrasing instincts to work with the MT output, which CSA Research’s practitioner surveys consistently associate with an average 20–30% higher fatigue rate per hour compared to conventional translation. A translator who works three hours of conventional translation in the morning and switches to MTPE in the afternoon is not starting the MTPE session at full speed.

Accounting for that overhead, the productivity improvement that actually holds hourly income is closer to 1.8× than 2.0×, because the lost output from the cognitive switching partially offsets the speed gain. The 1.8× threshold is editorial synthesis (source: our-estimate); no published study pins an exact ratio, but the break-even arithmetic is consistent with CSA Research’s fatigue-rate data.

What does 1.8× look like in practice?

Source: productivity estimates derived from TAUS Quality Dashboard post-editing studies and EUATC industry survey data. Your own figures will vary by CAT, TM density, and pair. Measure two MTPE sessions against two conventional sessions in the same week before accepting an ongoing MTPE arrangement.

The threshold test in plain terms: if the agency rate is below 60% of your full rate and you cannot demonstrate a personal post-editing output above 1.8× your conventional baseline, decline. State the reason: your measured productivity on this domain type does not clear the break-even ratio at the offered rate.

Named-tool comparison: DeepL Pro vs Claude/ChatGPT vs Trados Copilot

The tool decision is not which produces the best translation. It is which fits the workflow you are running inside a CAT.

Sources: DeepL Pro pricing as of May 2026; claude.ai/pricing; openai.com/pricing; RWS Trados product page; memoQ plans page.

The named recommendation: use DeepL Pro as your MTPE pre-editor inside the CAT. Use Claude or ChatGPT for three specific tasks only: terminology research before a project starts, definition lookups mid-session, and target-language final polish on segments where DeepL’s output is syntactically correct but tonally flat.

Do not use Claude or ChatGPT as a first-pass translation engine inside a CAT workflow. Neither has a native CAT plugin, which means copy-pasting segments, losing TM leverage, and destroying the consistency chain that CAT tools exist to maintain.

The CAT-tool integration question: why DeepL Pro wins on workflow, not output quality

Claude Sonnet and GPT-4o produce superior output on literary translation, legal nuance, and technical domains where sentence structure is complex. On a blind quality comparison, either model outperforms DeepL on EN → DE literary text by a margin that is visible to any competent reviewer.

The freelance translator’s workflow is not a blind quality comparison. It is a CAT session.

DeepL Pro’s plugin integration with Trados Studio, memoQ, Wordfast, and Phrase (SDL Trados plugin, memoQ server-side and desktop plugin, Wordfast MT integration) means that DeepL pre-populates segments inside the CAT, respecting TM leverage, segment order, and tag handling. The translator post-edits DeepL’s output against TM suggestions in one environment. Fuzzy matches from the TM take priority; DeepL fills the remainder. Consistency with previous deliverables is maintained automatically.

Recreating that workflow with Claude or ChatGPT requires: copying source segments out of the CAT, pasting them into the AI interface, copying the output back, checking tag preservation manually, and handling repetitions manually. The time cost of that friction exceeds the quality gain from the better model on everything except high-value literary or specialised legal translation where the per-word rate is already above 0.18 €/word.

At per-word rates above 0.18 €/word, the economics shift: the quality improvement Claude delivers on complex text is worth the workflow friction because the hourly income is already high enough that moderate productivity loss is acceptable. Below 0.18 €/word, the workflow integration matters more than the model quality.

ProZ.com’s rate calculator provides a useful cross-check on where your rates sit relative to the market; the tool is free and the underlying survey data is updated annually.

When to refuse MTPE work outright

Below 0.06 €/word source rate on MTPE, the math does not work at any realistic productivity ratio.

At 0.06 €/word MTPE against a 0.12 €/word full-translation baseline, the rate ratio is 50%. Achieving 48 €/hour at 0.06 €/word requires 800 words/hour post-editing output. Sustaining 800 words/hour across a full working day is not credible on general-domain text; the TAUS post-editing studies put average sustained post-editing output at 550–700 words/hour for experienced post-editors on high-quality MT.

Below 0.06 €/word, the answer to any MTPE offer is a written refusal with a rate floor statement. The format:

“My MTPE rate floor for [language pair] is [X] €/word, subject to TM analysis showing minimum 60% fuzzy-match density. I’m happy to quote this project at full translation rates.”

Agencies that want your pair and your quality will negotiate. Agencies that do not are not allocating work based on translator quality anyway; losing them does not lower your revenue per hour.

The second refusal trigger is low fuzzy-match density. An agency offering MTPE rates on a project with less than 50% TM match is asking you to translate from scratch while being paid for post-editing. Request the TM analysis before accepting any MTPE project. If the fuzzy density is below 50% on a general-domain job, re-quote at full translation rates.

For EN → ES below 0.08 €/word full rate: the MTPE rate band (0.03–0.07 €/word) is below 0.06 €/word on the lower end, which is the flat refusal threshold. If you are specialising in EN → ES at or below market rates, the MTPE offer trajectory is toward below-floor rates. The longer-term answer for EN → ES practitioners at the lower rate band is moving toward direct-client work, specialised domains (patent, pharmaceutical, financial), or higher-margin pairs.

Anti-patterns: what not to do

Using ChatGPT Free or Claude Free as a first-pass translation engine. The free tiers impose rate limits that break any sustained workflow. More importantly, pasting segments outside your CAT destroys TM leverage and tag consistency. If you are using a free AI tool for first-pass translation inside a professional MTPE workflow, you are doing more manual reconciliation than you are saving in draft time.

Running MT inside a CAT without measuring your own productivity. Most CAT tools can log words reviewed per hour; Trados Studio’s project reports and memoQ’s productivity statistics both surface this. If you are not measuring your post-editing output against your conventional translation output, you are accepting MTPE rates based on a productivity assumption that may not hold. Measure two sessions before agreeing to any ongoing MTPE arrangement.

Accepting agency MTPE rates without a written rate floor. Verbal rate agreements on MTPE are renegotiated downward every time MT quality improves. A written rate addendum stating your per-word floor and fuzzy-density minimum is the difference between a rate that holds and one that erodes quarterly.

Treating Claude or ChatGPT as a substitute for a terminology database. Both tools fabricate terminology in specialised domains. Claude Sonnet is significantly better than earlier models at acknowledging uncertainty, but it will still generate plausible-sounding technical terms that are not industry-standard. Use AI tools for terminology research to generate candidate terms and check definitions; verify against a domain-specific glossary, IATE (EU Terminology), or your own validated TB before including a term in a deliverable.

What changes this verdict

Cadence on this claim is 30 days. Three things would move it to Partial:

• DeepL releases a major quality uplift on EN → DE or EN → JA that closes the gap with Claude/GPT-4o on complex text. DeepL has shipped incremental quality improvements consistently; a step-change on literary or technical register would shift the named-tool recommendation toward a more conditional “it depends on domain” framing rather than a clear DeepL-first stance.
• A major CAT vendor (RWS/Trados, Kilgray/memoQ) ships a native Claude or OpenAI plugin with full segment and tag handling. This would change the CAT integration calculus entirely. Trados Copilot is RWS’s current answer; a direct Claude or GPT-4o Trados integration would be different.
• MTPE rate compression drops the agency standard below 35% of full rate across major pairs. CSA Research data shows rates have compressed 8–12% since 2021; a further step down in the 40–60% bracket would move the 1.8× break-even threshold higher, making viable MTPE increasingly rare at the freelance tier.

Status: Holding as of 12 May 2026. Next review: 11 Jun 2026.

Related operator reading

The solo-founder AI tool comparison that frames the Claude vs ChatGPT decision at a general level is at /operators/claude-pro-vs-chatgpt-plus-solo-founder/ (claim OPS-003). The AI delegation framework for 1-5 person businesses, including when AI fails on client-facing specialist work, is at /operators/what-to-delegate-to-ai/ (claim OPS-061).

Related enterprise reading

The procurement-side question of how enterprise language buyers are structuring MTPE vendor contracts is covered in /agentic-ai-readiness-diagnostic/ in the context of specialist-knowledge workflow automation.