Why I price budgets in blended cost
Provider price pages quote two numbers — input and output per million tokens — which makes budget math annoying. So for budget planning I collapse them into one: blended cost per million tokens, assuming a 3:1 input-to-output mix (75% of tokens in, 25% out), which is close to what I observe across chat, RAG, and coding workloads.
The formula is simple: blended = 0.75 x input price + 0.25 x output price. Claude Haiku 4.5 at $1 in / $5 out blends to $2 per million. Your $50 buys 25 million tokens.
If your workload is output-heavy (content generation) or input-heavy (document analysis), shift the ratio and the rankings move — Grok's cheap output makes it climb for generation work, while cheap-input models climb for RAG. The asterisk on the table is doing real work; the calculator lets you set your own ratio. For why the input-output distinction exists at all, see how AI API pricing works.
The formula is simple: blended = 0.75 x input price + 0.25 x output price. Claude Haiku 4.5 at $1 in / $5 out blends to $2 per million. Your $50 buys 25 million tokens.
If your workload is output-heavy (content generation) or input-heavy (document analysis), shift the ratio and the rankings move — Grok's cheap output makes it climb for generation work, while cheap-input models climb for RAG. The asterisk on the table is doing real work; the calculator lets you set your own ratio. For why the input-output distinction exists at all, see how AI API pricing works.
What those token numbers mean physically
Tokens are abstract until you convert them. Using the standard 0.75 words per token: 25M tokens (Haiku's $50 allotment) is about 18.8 million words — the complete works of Shakespeare twenty times over, or about 75,000 typical support conversations' worth of processing. Even Fable 5's 'small' 2.5M tokens is about 1.9 million words of combined reading and writing.
This is the most underrated fact in AI budgeting: for personal projects, prototypes, and small products, a $50/month budget on a sensibly chosen model is effectively unlimited. The conversion math is in tokens to dollars and how many tokens in 1,000 words.
This is the most underrated fact in AI budgeting: for personal projects, prototypes, and small products, a $50/month budget on a sensibly chosen model is effectively unlimited. The conversion math is in tokens to dollars and how many tokens in 1,000 words.
Three realistic $50 builds
Build one: a personal coding assistant. Heavy sessions burn 2M input + 150K output tokens; on Claude Sonnet 4.6 that's $8.25 a session, so $50 buys six serious sessions — or thirty-plus on Gemini 3.5 Flash. This is exactly why cascading matters at small scale.
Build two: a newsletter summarizer that ingests 200 articles a week (3,000 tokens each) and writes 200-token summaries. Weekly: 600K in, 40K out. Monthly: about 2.6M tokens — $5.20 on Haiku, $1.45 on Flash-Lite. You're using a tenth of the budget.
Build three: a niche RAG chatbot doing 5,000 questions a month (3,000 tokens of context in, 300 out each): 15M in + 1.5M out monthly. On Gemini 3.5 Flash that's $22.50 + $13.50 = $36 — fits with room to spare; on Sonnet it's $45 + $22.50 = $67.50 and you're over budget. At fixed budgets, the model choice IS the product decision.
Build two: a newsletter summarizer that ingests 200 articles a week (3,000 tokens each) and writes 200-token summaries. Weekly: 600K in, 40K out. Monthly: about 2.6M tokens — $5.20 on Haiku, $1.45 on Flash-Lite. You're using a tenth of the budget.
Build three: a niche RAG chatbot doing 5,000 questions a month (3,000 tokens of context in, 300 out each): 15M in + 1.5M out monthly. On Gemini 3.5 Flash that's $22.50 + $13.50 = $36 — fits with room to spare; on Sonnet it's $45 + $22.50 = $67.50 and you're over budget. At fixed budgets, the model choice IS the product decision.
Stretching the budget: the multipliers stack
Three discounts turn $50 into an effective $150-plus.
Prompt caching: chat and RAG workloads resend the same prefixes constantly; cached input bills at roughly a tenth of list rate. On input-heavy workloads (which is most of them — see the 3:1 mix) caching alone can nearly double your effective budget. Details in the caching guide.
Batch APIs: anything that can wait a few hours gets a flat 50% off on Anthropic and OpenAI. The summarizer in build two drops to $2.60/month.
Routing: send the easy 80% of requests to a budget model and reserve the mid tier for the hard 20%. A Flash-Lite-plus-Sonnet cascade typically lands at a quarter of all-Sonnet cost at near-identical perceived quality — the approach from the routing piece. Stack all three and the RAG bot from build three runs comfortably under $15.
Prompt caching: chat and RAG workloads resend the same prefixes constantly; cached input bills at roughly a tenth of list rate. On input-heavy workloads (which is most of them — see the 3:1 mix) caching alone can nearly double your effective budget. Details in the caching guide.
Batch APIs: anything that can wait a few hours gets a flat 50% off on Anthropic and OpenAI. The summarizer in build two drops to $2.60/month.
Routing: send the easy 80% of requests to a budget model and reserve the mid tier for the hard 20%. A Flash-Lite-plus-Sonnet cascade typically lands at a quarter of all-Sonnet cost at near-identical perceived quality — the approach from the routing piece. Stack all three and the RAG bot from build three runs comfortably under $15.
When $50 stops being enough
Budgets break at predictable thresholds, and knowing yours in advance beats discovering it on an invoice.
The math is linear: the RAG bot at $36/month for 5,000 questions becomes $360 at 50,000 and $3,600 at 500,000. The thresholds where teams typically re-architect: around $500/month is where caching and routing stop being optional; around $2,000/month is where serious eval-driven model selection pays for itself (a 20% saving covers real engineering time); around $10,000/month is where open-weight models on dedicated hosting enter the conversation, covered in the open-weight cost guide.
Set a billing alert at half your budget on day one. Token spend is silent — there's no progress bar — and every provider's console buries usage a few clicks deep. The diagnostic habit is in token usage auditing.
The math is linear: the RAG bot at $36/month for 5,000 questions becomes $360 at 50,000 and $3,600 at 500,000. The thresholds where teams typically re-architect: around $500/month is where caching and routing stop being optional; around $2,000/month is where serious eval-driven model selection pays for itself (a 20% saving covers real engineering time); around $10,000/month is where open-weight models on dedicated hosting enter the conversation, covered in the open-weight cost guide.
Set a billing alert at half your budget on day one. Token spend is silent — there's no progress bar — and every provider's console buries usage a few clicks deep. The diagnostic habit is in token usage auditing.