Reference
·Supporting source
Typhoon-2 Inference Cost Discount vs GPT-4o
~30-50% lower per-token cost (Thai-language workloads)
SCB 10X Typhoon-2 (released December 2025) reaches parity with GPT-4o on Thai-language tasks across M3Exam Thai, ThaiSum, and the Typhoon eval suite, at roughly 30 to 50 percent lower per-token inference cost when served from the AWS Bangkok region, Google Cloud Bangkok region, or NSTDA Lanta-class infrastructure. The cost differential plus latency advantage (no Singapore round-trip) flips the deployment decision for Thai-language regulated workloads from foundation-vendor wrappers to Thai LLM fine-tunes for net-new buys during 2026-27.
Figure in context
SCB 10X Typhoon-2 (released December 2025) reaches parity with GPT-4o on Thai-language tasks across M3Exam Thai, ThaiSum, and the Typhoon eval suite, at roughly 30 to 50 percent lower per-token inference cost when served from the AWS Bangkok region, Google Cloud Bangkok region, or NSTDA Lanta-class infrastructure. The cost differential plus latency advantage (no Singapore round-trip) flips the deployment decision for Thai-language regulated workloads from foundation-vendor wrappers to Thai LLM fine-tunes for net-new buys during 2026-27.
SCB 10X Typhoon-2 (released December 2025) reaches parity with GPT-4o on Thai-language tasks across M3Exam Thai, ThaiSum, and the Typhoon eval suite, at roughly 30 to 50 percent lower per-token inference cost when served from the AWS Bangkok region, Google Cloud Bangkok region, or NSTDA Lanta-class infrastructure. The cost differential plus latency advantage (no Singapore round-trip) flips the deployment decision for Thai-language regulated workloads from foundation-vendor wrappers to Thai LLM fine-tunes for net-new buys during 2026-27.
Time scope
Q4 2025 release; 2026 deployment
Source basis
Supporting source
Interpretation notes
What this tells you
SCB 10X Typhoon-2 (released December 2025) reaches parity with GPT-4o on Thai-language tasks across M3Exam Thai, ThaiSum, and the Typhoon eval suite, at roughly 30 to 50 percent lower per-token inference cost when served from the AWS Bangkok region, Google Cloud Bangkok region, or NSTDA Lanta-class infrastructure. The cost differential plus latency advantage (no Singapore round-trip) flips the deployment decision for Thai-language regulated workloads from foundation-vendor wrappers to Thai LLM fine-tunes for net-new buys during 2026-27.
What not to do with it
Use the linked report for interpretation and keep basis differences explicit.
Related figures
Adjacent numbers that add context without drowning the value.
Thai Enterprise AI Bilingual SaaS Run-Rate 2027
Operator disclosures, SET 56-1 One Report, PwC Thailand GenAI 2025, depa, NSTDA, Insight modelling
Thai Enterprise AI Bilingual SaaS Run-Rate 2026 Anchor
Operator disclosures, SET 56-1 One Report, PwC Thailand GenAI 2025, Insight triangulation
ThaiAIDeploy National AI Infrastructure Envelope
NSTDA Supercomputing announcements, depa AI Thailand Sandbox materials
Thai Financial-Services AI SaaS Run-Rate 2026
KBANK 56-1, SCB 56-1, Bluebik factsheet, Bitkub Capital Group AI disclosures
Thai Hospital and Contact-Centre AI PoC-to-Production Conversion
Bumrungrad 56-1, BDMS investor materials, Amity Solutions disclosures, Bangkok Post Inferense coverage
Thai Enterprise AI SaaS Operator Concentration 2026
Operator disclosures, SET 56-1 One Report, Insight share triangulation
Report context
Atlas actors in this figure's reports
Profiles covered in the report that cite this number.