Chọn model 101Choosing a model 101

Model đúng là model rẻ nhất vẫn vượt ngưỡng chất lượng trên dữ liệu của bạn. The right model is the cheapest one that still clears your quality bar on your own data.

Không phải model đứng đầu bảng xếp hạng. Đây là quyết định đầu tiên khi làm một app thật — và là quyết định nên đo lại mỗi khi có model mới. Not the one at the top of the leaderboard. It's the first decision in any real app — and one you should re-measure every time a new model ships.

Price bands per million output tokens, approximate (Claude prices checked October 2026), log scale: small $0.50–5, mid $9–15, large $12–50. Thinking pushes real cost further right. T1 $0.50–5 T2 $9–15 T3 $12–50 M1 thinking: token ẩn tính như output M1 thinking: hidden tokens billed as output $0.1 $1 $10 $100 giá mỗi 1 triệu token output, thang log price per 1M output tokens, log scale
Dải giá ước lượng theo hạng (giá Claude đối chiếu tháng 10/2026; hãng khác hãy xem trang giá của hãng). Mỗi vạch lớn là ×10: model nhỏ và model lớn cách nhau hai bậc độ lớn. Nét đứt là chi phí do model quyết định lúc chạy — bạn chỉ biết sau khi nhận hoá đơn. Approximate price bands by tier (Claude prices checked October 2026; for other vendors, check their pricing pages). Each major tick is ×10: small and large models sit two orders of magnitude apart. The dashed line is cost the model decides at runtime — you only see it on the bill.

Trước hết: có cần chọn model không?First: do you need to pick a model at all?

Trong nhiều trường hợp, câu trả lời đúng là không làm gì phức tạp.In many cases, the right answer is to not do anything elaborate.

Không dùng LLMSkip the LLM

Regex, rule, SQL hay thư viện có sẵn giải được thì dùng chúng: rẻ hơn, nhanh hơn, kết quả xác định.If regex, rules, SQL or an existing library solves it, use that: cheaper, faster, deterministic.

Dùng classifierUse a classifier

Nhãn cố định và đã có vài nghìn mẫu gán nhãn: classifier nhẹ trên embedding rẻ hơn hàng trăm lần mỗi request.Fixed labels and a few thousand labeled examples: a light classifier on embeddings is hundreds of times cheaper per request.

Không tối ưuDon't optimize

Vài trăm request mỗi ngày thì hoá đơn cỡ vài chục USD/tháng. Công so sánh model đắt hơn tiền tiết kiệm.A few hundred requests a day means a bill of tens of dollars a month. Comparing models costs more than it saves.

Phương án rẻ nhất khi buộc phải dùng LLM: một model hạng vừa của nhà cung cấp bạn đã có tài khoản, ghim đúng phiên bản, ship. Quay lại tối ưu khi hoá đơn hoặc chất lượng thật sự thành vấn đề — lúc đó bạn đã có log thật để dựng golden set.Cheapest option when you must use an LLM: a mid-tier model from a provider you already use, pinned to an exact version, shipped. Come back to optimize when cost or quality becomes a real problem — by then you'll have real logs for a golden set.

Chọn cái nào trong 30 giâyPick one in 30 seconds

Trả lời lần lượt. Dừng ở câu đầu tiên bạn trả lời "có".Answer in order. Stop at the first question you answer "yes" to.

  1. Giải được bằng rule, regex hay SQL?Solvable with rules, regex or SQL?

    Có: không dùng LLM.Yes: don't use an LLM.

  2. Luật hoặc hợp đồng cấm dữ liệu rời hạ tầng của bạn?Does law or contract forbid data leaving your infrastructure?

    Có: open-weight tự host A2. Nếu chỉ cần đúng vùng địa lý, xem bản API có data residency trước.Yes: self-hosted open-weight A2. If you only need a specific region, check APIs with data residency first.

  3. Dưới vài trăm request mỗi ngày?Under a few hundred requests a day?

    Có: model vừa T2 qua API A1. Dừng ở đây.Yes: mid-tier T2 via API A1. Stop here.

  4. Task cần suy luận nhiều bước?Does the task need multi-step reasoning?

    Có: model vừa + thinking T2 M1. Chưa đạt ngưỡng thì lên model lớn T3.Yes: mid-tier + thinking T2 M1. Below the bar? Step up to large T3.

  5. Chấm đúng/sai được bằng code?Can a script grade the output right or wrong?

    Có: thử model nhỏ + few-shot trước T1. Không: model vừa T2.Yes: try a small model + few-shot first T1. No: mid-tier T2.

Mọi nhánh: xác nhận trên golden set E1 trước khi chốt.Every branch: confirm on a golden set E1 before committing.

Bảy khái niệm, bốn nhómSeven concepts, four groups

Hạng và cách truy cập là hai trục độc lập: có model nhỏ đóng, có model lớn open-weight.Tier and access are independent axes: there are small closed models and large open-weight ones.

T: hạng model, to đến đâuT: tier, how big A: truy cập, chạy ở đâu và ai kiểm soátA: access, where it runs and who controls it M: chế độ, nghĩ bao lâuM: mode, how long it thinks E: đánh giá, biết mình chọn đúngE: evaluation, knowing you chose right
T1

Model nhỏSmall models

Hạng modelTier

Hạng rẻ nhất của mỗi hãng (nano, lite, haiku). Rất giỏi task hẹp có chỉ dẫn rõ.Each provider's cheapest tier (nano, lite, haiku). Very good at narrow, clearly specified tasks.

Dùng choUse for
Phân loại ticket, trích xuất JSON, router, kiểm duyệt lượt đầu.Ticket classification, JSON extraction, routing, first-pass moderation.
ĐượcGain
Rẻ hơn model lớn 10–100× trên output; TTFT thấp; chạy được trên mọi request.10–100× cheaper than large models on output; low TTFT; affordable on every request.
MấtLose
Sập ở task nhiều bước; bỏ sót chỉ dẫn dài; tiếng Việt tụt mạnh hơn.Breaks on multi-step tasks; drops long instructions; weaker on lower-resource languages.

Đổi khả năng xử lý đầu vào mơ hồ lấy giá và tốc độ đủ thấp để chạy trên mọi request.Trades handling of ambiguous input for a price and speed low enough to run on every request.

Dấu hiệu: bạn viết được hàm chấm đúng/sai bằng code.Tell-tale sign: you can write a function that grades the output.

T2

Model vừaMid-tier models

Hạng modelTier

Hạng giữa. Điểm khởi đầu hợp lý cho đa số app, thường có tuỳ chọn bật thinking.The middle tier. The sensible starting point for most apps, usually with an optional thinking mode.

Dùng choUse for
Chatbot RAG, agent vài bước gọi tool, code hằng ngày, tóm tắt tài liệu dài.RAG chatbots, few-step tool agents, everyday coding, long-document summaries.
ĐượcGain
Phần lớn chất lượng model lớn với giá thấp hơn 2–5×; hiếm khi quá tệ, hiếm khi quá đắt.Most of a large model's quality at 2–5× lower cost; rarely terrible, rarely too expensive.
MấtLose
Không rẻ nhất cho task đơn giản; thua rõ ở agent chạy dài; dễ bị giữ mãi không ai đo lại.Not cheapest for simple tasks; clearly behind on long agent runs; tends to stay unmeasured forever.

Đổi phần chất lượng đỉnh và phần tiết kiệm tối đa lấy một điểm khởi đầu ít sai nhất.Trades peak quality and maximum savings for the starting point least likely to be wrong.

Dấu hiệu: cần hiểu ngữ cảnh, nhưng hiếm khi quá năm bước suy luận.Tell-tale sign: needs context, but rarely more than five reasoning steps.

T3

Model lớnLarge models

Hạng modelTier

Model mạnh nhất mỗi hãng bán. Suy luận mạnh, context dài, chạy agent hàng trăm bước.The strongest model each provider sells. Strong reasoning, long context, agents that run hundreds of steps.

Dùng choUse for
Agent code nhiều file, phân tích pháp lý/tài chính, nghiên cứu nhiều bước, làm giám khảo.Multi-file coding agents, legal/financial analysis, multi-step research, acting as judge.
ĐượcGain
Chất lượng cao nhất; xử lý input mơ hồ; đôi khi rẻ hơn tính theo task vì ít retry.Highest quality; handles vague input; sometimes cheaper per task thanks to fewer retries.
MấtLose
Đắt gấp vài lần T2, hàng chục lần T1; chậm hơn; rate limit chặt hơn.Several times T2, tens of times T1; slower; tighter rate limits.

Đổi tiền và độ trễ lấy xác suất làm đúng ngay lần đầu.Trades money and latency for the probability of getting it right the first time.

Dấu hiệu: một lần sai đắt hơn nhiều lần chi phí một lần gọi.Tell-tale sign: one mistake costs far more than one call.

A1

Model đóng (API)Closed models (API)

Truy cậpAccess

Không công bố trọng số. Gọi qua API, trả theo token, nhà cung cấp lo hạ tầng.Weights not published. Called via API, paid per token, infrastructure handled by the provider.

Dùng choUse for
Ship trong vài tuần, lưu lượng thất thường, cần chất lượng đỉnh, team không có MLOps.Shipping in weeks, bursty traffic, top quality needed, no MLOps on the team.
ĐượcGain
Không vận hành gì; lưu lượng bằng 0 thì chi phí bằng 0; cache read giảm từ 90% trở lên (tuỳ model), batch ~50%.Nothing to operate; zero traffic, zero cost; cache reads 90% off or more (depends on model), batch ~50% off.
MấtLose
Dữ liệu rời hạ tầng; model có thể bị deprecate (ghim model ID, có fallback sang hãng thứ hai); sự cố của họ thành sự cố của bạn.Data leaves your infra; models get deprecated (pin the model ID, keep a fallback to a second provider); their outages become yours.

Đổi quyền kiểm soát dữ liệu, phiên bản và lịch nâng cấp lấy tốc độ ra sản phẩm và chất lượng đỉnh.Trades control over data, versions and upgrade timing for speed to market and top quality.

Dấu hiệu: không ai trong team muốn trực GPU lúc 3 giờ sáng.Tell-tale sign: nobody wants to be on call for GPUs at 3 a.m.

A2

Open-weight

Truy cậpAccess

Trọng số tải về được. Chạy qua host serverless hoặc tự host. Open-weight không đồng nghĩa open source.Downloadable weights. Run through a serverless host or self-host. Open-weight is not open source.

Dùng choUse for
Dữ liệu không được rời mạng nội bộ, lưu lượng lớn và đều, fine-tune sâu, on-device.Data that can't leave your network, large steady traffic, deep fine-tuning, on-device.
ĐượcGain
Kiểm soát hoàn toàn; bản host sẵn có thể rất rẻ (DeepSeek V4.1-Flash: $0.15–0.30 / $0.60–1.20 mỗi 1M token, 10/2026 — xem trang giá của hãng); không ai deprecate.Full control; hosted prices can be very low (DeepSeek V4.1-Flash: $0.15–0.30 / $0.60–1.20 per 1M tokens, Oct 2026 — check the vendor page); nobody deprecates it.
MấtLose
Tự gánh vận hành; GPU nhàn rỗi vẫn tốn tiền; tầng đỉnh vẫn kém model đóng.You carry operations; idle GPUs still bill; still behind closed models at the very top.

Đổi công vận hành lấy quyền kiểm soát dữ liệu, phiên bản và chi phí biên.Trades operational effort for control over data, versions and marginal cost.

Dấu hiệu: có điều khoản pháp lý ghi rõ dữ liệu phải nằm ở đâu.Tell-tale sign: a legal clause spells out where the data must live.

M1

Reasoning / thinkingReasoning / thinking

Chế độ chạyRun mode

Model nháp, thử và sửa trước khi trả lời. Phần lớn model hiện nay bật/tắt hoặc chỉnh effort được theo request.The model drafts, tests and corrects before answering. Most current models let you toggle it or set effort per request.

Dùng choUse for
Toán, logic, debug khó, lập kế hoạch cho agent, rà mâu thuẫn trong hợp đồng.Math, logic, hard debugging, agent planning, finding contradictions in contracts.
ĐượcGain
Chính xác hơn hẳn ở task nhiều bước; model vừa + thinking có thể thay model lớn.Much more accurate on multi-step tasks; mid-tier + thinking can replace a large model.
MấtLose
Token ẩn tính như output, chi phí khó đoán; độ trễ lên hàng chục giây; hại task đơn giản.Hidden tokens billed as output, cost unpredictable; latency in tens of seconds; hurts simple tasks.

Đổi độ trễ và một hoá đơn khó đoán lấy độ chính xác ở bài toán nhiều bước.Trades latency and an unpredictable bill for accuracy on multi-step problems.

Dấu hiệu: người giỏi phải nháp ra giấy mới giải được.Tell-tale sign: a skilled person would need scratch paper.

E1

Đánh giá trên golden setGolden-set evaluation

Đánh giáEvaluation

50–200 input thật của app, có đáp án hoặc rubric. Chạy 2–3 model, chọn model rẻ nhất vượt ngưỡng đã đặt trước.50–200 real inputs from your app, with answers or a rubric. Run 2–3 models; pick the cheapest that clears a bar set in advance.

Dùng choUse for
Chọn model lần đầu, thử model mới. Cách dựng và chấm golden set: Evals 101.First model choice, trying new models. Building and grading the golden set: Evals 101.
ĐượcGain
Số đo trên chính dữ liệu của bạn; có model mới thì chạy lại trong vài phút.Measurements on your own data; rerun in minutes when a new model ships.
MấtLose
Output không xác định (model Claude hiện hành không cho chỉnh temperature) nên chạy mỗi mẫu 3 lần; 100 mẫu ở mức ~80% có sai số ±8 điểm.Outputs aren't deterministic (current Claude models don't let you set temperature), so run each sample 3 times; 100 samples at ~80% carry ±8 points of error.

Đổi vài ngày công lấy khả năng chọn model bằng số đo thay vì cảm giác.Trades a few person-days for choosing models by measurement instead of by feel.

Dấu hiệu: bạn sắp thay đổi bất cứ thứ gì ảnh hưởng tới output.Tell-tale sign: you're about to change anything that affects output.

Bảng đánh đổiTrade-off table

E1 không nằm trong bảng: nó là công cụ để đo các trục này, không phải một lựa chọn trên chúng.E1 isn't in the table: it's the tool that measures these axes, not an option on them.

Chất lượngQuality Chi phí/taskCost per task Độ trễLatency Công vận hànhOps effort Kiểm soát dữ liệuData control
T1NhỏSmall Thấp–vừa, tốt ở task hẹpLow–medium, strong on narrow tasks Rất thấpVery low Rất thấpVery low ThấpLow Tuỳ cách truy cậpDepends on access
T2VừaMid CaoHigh VừaMedium ThấpLow ThấpLow Tuỳ cách truy cậpDepends on access
T3LớnLarge Cao nhấtHighest CaoHigh Vừa–caoMedium–high ThấpLow Tuỳ cách truy cậpDepends on access
A1ĐóngClosed Đỉnh thị trường ở đâyMarket's top is here Theo token, không phí cố địnhPer token, no fixed cost Phụ thuộc nhà cung cấpProvider-dependent Rất thấpVery low ThấpLow
A2Open-weight Gần đỉnh, chưa bằngNear the top, not quite Rất thấp nếu volume đều; cao nếu GPU nhàn rỗiVery low with steady volume; high with idle GPUs Tự kiểm soátUnder your control Cao (tự host)High (self-host) Cao nhấtHighest
M1Thinking Tăng mạnh ở task nhiều bướcLarge gains on multi-step tasks Tăng nhiều lần, khó đoánRises several-fold, unpredictable Tăng mạnhRises sharply ThấpLow Không đổiUnchanged
Hạng (ước lượng, mỗi 1M token)Tier (approx., per 1M tokens)InputOutput
T1~$0.10–1~$0.50–5
T2~$1.5–3~$9–15
T3~$2–10~$12–50

Giảm phổ biến: cache read tốn 10% giá input hoặc ít hơn, tuỳ model (Opus 5.5 và Sonnet 5.5 là 5%, Fable 5.1 là 2,5%); batch API ~50%. Giá đổi vài tháng một lần — dùng để hiểu tỷ lệ, không để chốt ngân sách.Common discounts: cache reads at 10% of input price or less, depending on model (Opus 5.5 and Sonnet 5.5 5%, Fable 5.1 2.5%); batch APIs ~50% off. Prices change every few months — use them for ratios, not budgets.

Anti-patterns

Sai lầmMistake Vì sao saiWhy it's wrong Sửa thế nàoFix
Chọn theo bảng xếp hạng benchmarkChoosing by leaderboard rankĐo dữ liệu của người khácIt measures someone else's dataChạy 2–3 model trên golden set của mìnhRun 2–3 models on your own golden set
Model lớn cho mọi requestLarge model for every requestTrả gấp hàng chục lần cho task model nhỏ làm đượcPaying tens of times more for tasks a small model can doRouter: nhỏ trước, escalate khi validate failRouter: small first, escalate on validation failure
So giá theo tokenComparing price per tokenBỏ qua token suy luận, retry, tokenizer khác nhauIgnores reasoning tokens, retries, tokenizer differencesTính chi phí mỗi task trên workload thậtCompute cost per task on the real workload
Bật thinking cho mọi requestThinking on for every requestTrễ và đắt cả ở task không cầnSlower and pricier where it isn't neededBật theo loại task, effort thấp làm mặc địnhEnable per task type, low effort by default
Tự host để "tiết kiệm" khi volume thấpSelf-hosting "to save money" at low volumeGPU nhàn rỗi vẫn tính tiền theo giờIdle GPUs still bill by the hourHost serverless tới khi lưu lượng lớn và đềuServerless host until traffic is large and steady
Đổi model không chạy lại evalSwitching models without rerunning evalsHồi quy âm thầm, người dùng phát hiện trướcSilent regressions users find firstĐưa golden set vào CIPut the golden set in CI
Chỉ nhìn độ trễ trung bìnhWatching only average latencyp95 mới là trải nghiệm tệ nhất người dùng gặpp95 is the worst experience users actually getĐo và đặt SLA trên p50 và p95Measure and set SLAs on p50 and p95

Ba câu cần nhớThree things to remember