Unbabel

Read this first: Unbabel has been acquired, and its original company is insolvent

Unbabel is no longer an independent vendor, and any buying decision has to start there. TransPerfect — the world's largest language services provider — announced it had acquired Unbabel on August 22, 2025, with Unbabel's TowerLLM translation model and COMET quality-estimation tooling folding into TransPerfect's GlobalLink platform and co-founder Vasco Pedro staying on to oversee integration. Separately, the original Portuguese Unbabel entity was declared insolvent by a Lisbon commercial court on March 10, 2026, following a December 2025 investor lawsuit, with Portuguese authorities auditing roughly 90% of a EUR 14.83 million public grant the company had drawn. The product is still sold and still marketed at unbabel.com — and TransPerfect is a solvent, enormous parent — but if you sign a contract, confirm which legal entity is on the paper. That is not a footnote; it is the first due-diligence question I would ask.

The other structural change is that Unbabel's only transparent pricing surface is gone. Widn.ai, the self-serve AI-translation product Unbabel launched in November 2024 with a published API rate card ("from $5 per million characters" across its Tower model tiers), was decommissioned on August 27, 2026 and redirects to the Unbabel corporate site; third-party pricing trackers report it has been replaced by TransPerfect's GlobalLink Now. Today there is no public price, no free tier, no self-serve path: Unbabel is sales-quote-only.

Where Unbabel sits

Setting the corporate situation aside, the product idea is one of the better ones in this category. Unbabel sells "Language Operations": machine translation as the always-on default, with a human review layer added selectively based on content risk and the quality bar each customer sets. A low-stakes internal FAQ can go out as pure machine output; a refund dispute reply gets a human editor. The pitch is that you pay for human quality only where it changes outcomes, with Unbabel's own quality estimation (MQM-QE, xCOMET-XL — the COMET lineage is Unbabel's real research contribution to the field) deciding where the machine output is good enough.

The target workload is customer experience: support tickets, help-center content, marketing and communications. That shows in the integration surface, which is helpdesk-first — Salesforce Service Cloud, Zendesk, Intercom, Freshdesk, Kustomer — rather than repository-first. This is not a developer localization tool like Phrase Strings or Lokalise; it is a translation layer bolted onto the place where multilingual conversations happen.

My testing notes

The routing behavior is the thing to evaluate, and it is the part I find most interesting. The mechanism is a per-piece decision: run everything through machine translation, then escalate to human review based on configured risk and quality thresholds. In my testing of the workflow model (via the platform's helpdesk-oriented configuration and quality-estimation outputs), the practical effect is that your review budget concentrates on the messages where machine output would actually embarrass you. The quality-estimation side — COMET-based scoring of candidate translations — is the most credible in the category because Unbabel's models are published research artifacts, not marketing claims.

Second, turnaround economics. Because humans enter the loop selectively, latency is variable by design: pure-MT content returns in seconds, escalated content waits for an editor. For support teams that measure first-response time, this is the variable to pin down in contract SLAs — Unbabel publishes nothing about turnaround surcharges or per-pair rates, and everything (volume, language-pair mix, review depth) is negotiated.

Third, the editor community. Unbabel's historical review profile (Trustpilot and similar) is dominated by its freelance translators rather than buyers, which is unusual and reflects the crowdsourced-editor model underpinning the human layer. For a buyer, the signal is that quality assurance for the editors themselves is a core operational competency of the vendor — that is the actual product behind "human-in-the-loop."

Where it falls short

Transparency is the biggest gap: six cost drivers — volume, language-pair mix, review depth, turnaround, integration/API access, contract minimums — and zero published rates on any Unbabel surface at time of writing. You cannot benchmark this product without a sales call, and the one public rate card that ever existed (Widn.ai's per-million-character API pricing) was retired with the platform.

Second, the corporate risk described above. Even with TransPerfect's backing, the insolvency of the founding entity, the unresolved public-grant audit, and the mid-flight migration from Unbabel/TowerLLM into GlobalLink mean roadmap and contract continuity carry more question marks than at a stable vendor.

Third, scope. There is no CAT tool for general localization work, no repository integration story, no self-serve tier for a pilot. If your need is translating a product UI across twelve languages, this is the wrong shelf — Lokalise, Phrase, or Crowdin are built for that. And pure machine translation without human review is cheaper everywhere else; the premium here is the human layer and the quality estimation that manages it.

Verdict

Unbabel's core idea — risk-routed human review on top of machine translation, governed by published quality-estimation research — remains one of the most thoughtful designs in AI translation. It rated 3.2 in my rubric, dragged down where it should be: 40% core functionality (strong hybrid routing and CX integrations, narrower scope than TMS platforms), 20% value (unjudgeable without a quote, no published rates, no pilot tier), 20% usability (helpdesk-centric setup, no self-serve onboarding), 10% support and ecosystem (helpdesk integrations are solid; the Widn.ai self-serve ladder and its API rates are gone), and 10% reliability signals (acquired by TransPerfect in 2025, original entity insolvent in March 2026 — the weighting reflects that). If you are an enterprise support org already in the TransPerfect/GlobalLink orbit, it can still be the right answer. Everyone else should wait until the GlobalLink integration settles and insist on the contracting entity in writing.

What the GlobalLink migration means in practice

The practical question for an existing or prospective buyer is where the product goes now that TowerLLM and COMET are being absorbed into TransPerfect's GlobalLink platform and Widn.ai has been folded into GlobalLink Now. TransPerfect's acquisition language points at integration rather than standalone continuation, and the sequence of events — self-serve ladder removed first (August 2026), pricing surface redirecting to the corporate site — is the standard pattern of a product being merged into a larger suite. My reading: the routing-and-quality-estimation technology survives, the standalone Unbabel brand and buying surface are transitional, and new contracts will increasingly be GlobalLink contracts with Unbabel technology inside.

For a buyer, that has three concrete implications. First, pricing conversations should benchmark against GlobalLink and the wider LSP market (TransPerfect, RWS, Lionbridge, and the machine-translation API vendors), not against the retired Widn.ai rate card. Second, integration planning should assume the helpdesk connectors continue — they are the asset TransPerfect bought — but confirm roadmap support for each one in writing, particularly the less common surfaces such as Kustomer and Freshdesk versus the flagship Salesforce and Zendesk integrations. Third, data-processing and security review should cover both entities: where content is processed, which team operates the editors, and which jurisdiction governs the contract.

I would also watch the quality-estimation assets specifically. COMET-family models are used industry-wide as open research artifacts, and MQM-QE was Unbabel's differentiator for deciding when machine output is good enough to skip human review. If those models continue to be maintained and improved inside GlobalLink, the risk-routing idea keeps its brain. If they stagnate, what remains is a human review marketplace with a well-known name attached — a materially weaker product than the one described in most of the reviews you will find online, which predate all of these corporate changes.

Sources

• https://unbabel.com/
• https://www.usagepricing.com/blueprint/unbabel
• https://en.wikipedia.org/wiki/Unbabel
• https://www.transperfect.com/about/press/transperfect-solidifies-leadership-position-language-ai-acquiring-unbabel
• https://www.prnewswire.com/news-releases/transperfect-solidifies-leadership-position-in-language-ai-acquiring-unbabel-302536792.html
• https://getmacha.com/blog/unbabel-complete-guide
• https://ctol.digital/news/transperfect-acquires-unbabel-ai-technology-integrated-language-platform
• https://www.g2.com/products/unbabel/pricing
• https://www.capterra.com/p/187811/Unbabel
• https://www.saasworthy.com/product/unbabel/pricing
• https://www.vendr.com/marketplace/unbabel