DevTk.AI
AI Model SelectionCoding AgentsModel RoutingAPI PricingGPT-6.2Claude Opus 5.5Gemini 4

Which AI Model for Which Job (November 2026): Planner, Coder, Long Runs and Bulk Workers, with Prices

A role-by-role model picker for coding agents: GPT-6.2 Astra or Gemini 4 Neon to plan, GPT-6.2 Sol, Sonnet 5.5 or Gemini 4 Hydrogen to implement, Opus 5.5 for long runs, GPT-6 Luna or Haiku 5.5 for bulk work. Prices, context, budget picks and monthly cost math.

By Published 17 min read

Short answer: don’t run every step of a coding agent on a flagship model. Give planning and the hardest calls to a top model, everyday edits to a mid-tier model, and repetitive sub-agent work to a cheap one. In the worked example below, that turns a $588/month all-GPT-6.2-Astra bill into $149 with OpenAI models only, $120 with Anthropic models only, or $22–$46 with a budget mix.

How the picks were made:

  • Role fit comes from each vendor’s own positioning for the model (how the provider describes what it is for), not from benchmarks. We don’t publish benchmark claims.
  • Prices are standard list prices per 1M tokens from the DevTk.AI model database, checked against each provider’s pricing page in mid-November 2026. Each model name links to its page with cache, batch and long-context details.
  • Real cost per task depends on how many tokens a model uses. Reasoning tokens bill as output, and a model that needs two attempts costs twice as much. A cheaper per-token model is not automatically cheaper per finished task, so test the cheaper tier first and escalate only what fails (more on that below).

Quick Answer: One Model per Role

RolePickInput / output per 1MContextWhy (vendor’s positioning)
Planner / architect / reviewerGPT-6.2 Astra$8 / $401.05MOpenAI’s flagship, “our most capable model for your hardest projects”
Implementer (everyday coding)GPT-6.2 Sol$2 / $101.05MOpenAI’s “go-to model for everyday work, from coding to powering agents at scale”
Long autonomous runsClaude Opus 5.5$4 / $201MAnthropic’s current default model for long-running agentic coding and knowledge work
Bulk / sub-agent workersGPT-6 Luna$0.10 / $0.501.05MOpenAI: for focused, repeatable, high-volume work
Classification, extraction, routingClaude Haiku 5.5$0.10 / $0.50 (prompts ≤100K)1MAnthropic’s fastest current model, for classification, extraction and routing
Screenshots, computer use, writingGPT-6.1 Terra$1 / $51.05MOpenAI: visual reasoning, computer use and natural writing, at half Sol’s price
Autocomplete / fill-in-the-middleCodestral 25.08$0.30 / $0.90128KMistral’s low-latency code model for completion and FIM
Local / self-hostedMistral Large 4$0.70 / $2.10 (API)1MMistral’s open-weight flagship; call the API or run the weights yourself

Every role has strong alternatives from other vendors. The sections below list them with the same price and positioning columns.

Planner / Architect

The planner reads the task and the codebase, writes the plan, splits it into steps and reviews the result. It handles a small share of the tokens but shapes everything downstream, so this is where a top model earns its price.

ModelInput / outputCached inputContextVendor positioning
GPT-6.2 Astra$8 / $40$0.401.05MOpenAI’s most capable model for the hardest projects (released 2026-11-17)
Gemini 4 Neon$3 / $15$0.151MGoogle: complex software engineering, knowledge work and science workflows
Gemini 4 Carbon$2 / $10 to Nov 30, then $4 / $20$0.10, then $0.201MGoogle’s most capable Gemini reasoning model: complex software engineering, math and science, autonomous multi-step agents
Claude Fable 5.2$8 / $40$0.201MSuccessor to Fable 5.1, Anthropic’s highest-capability broadly available model for long-horizon agentic work, at 20% lower rates
Qwen4 Max$1.50 / $4.50$0.0751MAlibaba’s Qwen4 flagship for complex reasoning and coding

Notes:

  • Gemini 4 Carbon’s price doubles on December 1, 2026. From then on, Gemini 4 Neon ($3 / $15) is the cheaper of the two. Neither is on the Gemini API free tier.
  • GPT-6.2 prompts above 272K input tokens bill the whole request at long-context rates ($16 / $60 for Astra). A planner that loads a large repo can cross that line; trimming context back under 272K halves the input price.
  • Gemini 4 Neon bills thinking tokens as output, like the other reasoning models here.

For the full GPT-6.2 price sheet (cache writes, Batch, Fast and Ultrafast), see GPT-6.2 Astra and Sol pricing.

Implementer (Everyday Coding)

The implementer turns the plan into edits, runs tools and fixes tests. It carries most of the volume, so the per-token price matters a lot here.

ModelInput / outputCached inputContextVendor positioning
GPT-6.2 Sol$2 / $10$0.101.05MOpenAI’s go-to model for everyday work, from coding to agents at scale
Claude Sonnet 5.5$2 / $10$0.101MAnthropic’s current balanced Claude model
Gemini 4 Hydrogen$1.50 / $7.50$0.0751MGoogle’s stable Gemini 4 model for everyday coding and agentic workflows
Kimi K3.1 Pro$2 / $10$0.201MMoonshot AI’s newest Kimi flagship, a third cheaper than Kimi K3
Grok 4.7$2 / $6$0.50500KxAI’s current frontier model for coding, agents and knowledge work

Per token, GPT-6.2 Sol and Claude Sonnet 5.5 cost exactly the same, including the $0.10 cache read. Gemini 4 Hydrogen is 25% cheaper on both input and output, and it is available on the Gemini API free tier, so it is the easiest one to trial. Since the list prices are this close, pick the implementer that finishes your tasks in fewer tokens and fewer retries, and measure that on your own repo.

If you are on GPT-6.1 Sol, GPT-6.2 Sol costs the same, so new projects should start on 6.2.

Long Autonomous Runs

Long runs (an agent working through a backlog for an hour or more) resend a growing prompt on every turn. That makes the cached-input price and the long-context rules matter more than the headline input price.

ModelInput / outputCached inputContextVendor positioning
Claude Opus 5.5$4 / $20$0.201MAnthropic’s current default for long-running agentic coding and knowledge work
Claude Fable 5.2$8 / $40$0.201MFable 5.1’s successor for long-horizon agentic work; full 1M context at standard rates
GLM-5.3$1.40 / $4.40$0.261MZ.AI’s GLM flagship for complex coding and long-horizon agents; reasoning always on
Gemini 3.8 Flash$0.75 / $3.75 to Dec 31, then $1.50 / $7.50$0.075, then $0.151MGoogle’s GA Flash flagship for long-horizon coding, autonomous agents and enterprise workflows
Muse Wonder 1.0 (preview)$4 / $20$0.201MMeta’s most capable Muse model, for complex multi-file software engineering and long-horizon agent work

Prompt size changes the bill differently by vendor:

  • OpenAI GPT-6.2 / GPT-6.1: above 272K input tokens the entire request bills at long-context rates (GPT-6.2 Sol goes from $2 / $10 to $4 / $15).
  • xAI Grok 4.7: long-context rates ($4 / $12) apply to all tokens once the prompt reaches 200K.
  • Anthropic Sonnet 5.5 and Fable 5.2: the full 1M context is billed at standard rates. Our data lists no long-context tier for Opus 5.5 either.
  • Gemini 3.8 Flash and Muse Wonder 1.0: no long-context tier in our data, but Gemini 3.8 Flash’s introductory price ends on December 31, 2026.

Bulk Workers and Sub-Agents

Sub-agents that search files, summarize logs, label data, validate JSON or route requests do the same narrow job thousands of times. This is where the cheapest tier pays off, and where a model’s prompt-size tiers matter most.

ModelInput / outputCached inputContextVendor positioning / caveat
GPT-6 Luna$0.10 / $0.50$0.011.05MOpenAI: focused, repeatable, high-volume work. Above 272K input: $0.20 / $0.75
Claude Haiku 5.5$0.10 / $0.50$0.011MAnthropic: classification, extraction and routing. Prompts above 100K: $0.50 / $2.50
Muse Ember 1.0$0.10 / $0.50$0.0041MMeta: high-volume sub-agents, tool calls, routing and extraction. Above 200K input: $0.30 / $1.50
DeepSeek V4.1 Flash$0.30 / $1.20 peak, $0.15 / $0.60 off-peak$0.006 peak, $0.003 off-peak1MPicked on price: very cheap cache hits for loops that resend the same repo context; 384K max output
Qwen3.7 Flash$0.03 / $0.13 (≤32K)$0.0061MAlibaba: multimodal agents and coding. 32K–256K: $0.10 / $0.40; 256K–1M: $0.20 / $0.80
GLM-5.3 Flash$0.15 / $0.50$0.031MZ.AI’s lowest-cost GLM-5.3 route
Mistral Small 4$0.15 / $0.60$0.015256KMistral’s low-cost model for high-volume production routing

Keep sub-agent prompts small. That is the point of a sub-agent, and it also keeps you under the price jumps: Haiku 5.5 costs 5x more once a prompt goes over 100K tokens, and Qwen3.7 Flash’s cheapest rate only covers prompts up to 32K. DeepSeek’s peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; every other hour, weekends included, is half price.

Vision, Autocomplete and Local Models

  • Screenshots and computer use: GPT-6.1 Terra ($1 / $5) is the model OpenAI positions for visual reasoning, computer use and natural writing. For high-throughput multimodal jobs on a budget, Gemini 3.5 Flash-Lite ($0.30 / $2.50) is Google’s low-cost GA option.
  • Autocomplete: Codestral 25.08 ($0.30 / $0.90, 128K) is built for completion and fill-in-the-middle, a different job from agentic coding.
  • Local / open weights: in our data, Mistral Large 4 is open-weight (52B active / 1.05T total parameters), the Ministral 3 models (3B, 8B, 14B) are small Apache 2.0 open-weight models, and Moonshot publishes full weights for Kimi K3. Running them yourself trades token bills for hardware; the VRAM calculator and the self-hosting vs API cost guide cover that math.

Budget Version: Cheapest Capable Models

Same filter as the “cheapest model that can still do the job” list on our homepage: tool calling, JSON/structured output, at least 200K context, released in the last nine months, one model per mainstream provider, sorted by blended price (3 input : 1 output).

ModelProviderInput / outputCached inputContextWatch out for
Qwen3.7 FlashAlibaba$0.03 / $0.13$0.0061MPrice applies to prompts ≤32K
GPT-6 LunaOpenAI$0.10 / $0.50$0.011.05MAbove 272K input: $0.20 / $0.75
Claude Haiku 5.5Anthropic$0.10 / $0.50$0.011MAbove 100K input: $0.50 / $2.50
GLM-5.3 FlashZ.AI$0.15 / $0.50$0.031M—
Mistral Small 4Mistral$0.15 / $0.60$0.015256K32K max output
DeepSeek V4.1 FlashDeepSeek$0.30 / $1.20$0.0061MPeak rate; off-peak is half
Gemini 3.5 Flash-LiteGoogle$0.30 / $2.50$0.031MOutput-heavy jobs cost more
Grok Build 0.1xAI$1 / $2$0.20256KPublic beta

And the role table again, with the cheapest model whose vendor positioning fits each role:

RoleBudget pickInput / outputContextVendor positioning
PlannerQwen4 Max$1.50 / $4.501MQwen4 flagship for complex reasoning and coding
ImplementerGemini 3.8 Flash$0.75 / $3.75 (to Dec 31)1MLong-horizon coding and autonomous agents
Long runsGLM-5.3$1.40 / $4.401MComplex coding and long-horizon agents
Cache-heavy loopsDeepSeek V4.1 Flash$0.30 / $1.20 peak1MPicked on price ($0.006 cache hits)
Bulk workersGPT-6 Luna or Claude Haiku 5.5$0.10 / $0.501M+Repeatable high-volume work; classification, extraction, routing
Short-prompt bulkQwen3.7 Flash$0.03 / $0.13 (≤32K)1MMultimodal agents and coding

Worked Example: Planner + Workers vs One Flagship

Example workload: a coding agent that sends 100M input tokens a month, 70% of them cache hits, and receives 8M output tokens. In the split setups the planner handles 10% of the traffic, the implementer 60% and bulk workers 30%. Standard mode, prompts under each model’s long-context threshold (100K for Haiku 5.5, 272K for OpenAI), DeepSeek at peak rates, cache-write charges left out for simplicity.

Each share is priced the same way: uncached input × input price + cached input × cached price + output × output price. For the OpenAI split:

Role (share)ModelUncached inputCached inputOutputSubtotal
Planner (10%)GPT-6.2 Astra3M × $8 = $24.007M × $0.40 = $2.800.8M × $40 = $32.00$58.80
Implementer (60%)GPT-6.2 Sol18M × $2 = $36.0042M × $0.10 = $4.204.8M × $10 = $48.00$88.20
Bulk (30%)GPT-6 Luna9M × $0.10 = $0.9021M × $0.01 = $0.212.4M × $0.50 = $1.20$2.31
Total$149.31

All setups, same workload:

SetupPlanner (10%)Implementer (60%)Bulk (30%)Monthly total
Everything on GPT-6.2 Astra$588.00
Everything on Claude Opus 5.5$294.00
Everything on GPT-6.2 Sol$147.00
OpenAI split: Astra / Sol / Luna$58.80$88.20$2.31$149.31
Anthropic split: Opus 5.5 / Sonnet 5.5 / Haiku 5.5$29.40$88.20$2.31$119.91
Budget split: Qwen4 Max / Gemini 3.8 Flash / GPT-6 Luna$8.63$34.65$2.31$45.59
Budget split with DeepSeek V4.1 Flash (peak) as implementer$8.63$11.41$2.31$22.35

The single-model rows are the same math on all 100M / 8M tokens. For example, everything on Astra is 30M × $8 + 70M × $0.40 + 8M × $40 = $240 + $28 + $320 = $588.

What the numbers say:

  • The split is about 75% cheaper than running everything on the flagship ($149.31 vs $588). Against all-Opus 5.5, the Anthropic split saves about 59% ($119.91 vs $294).
  • The split is not cheaper than running everything on the mid-tier model. All-Sol costs $147.00. Moving the planner from Sol to Astra adds $44.10, and moving bulk work from Sol to Luna saves $41.79, so the split costs $2.31 more. What you get for that is your hardest 10% of calls on the flagship. If your mid-tier model already plans well enough, the cheaper move is all-mid-tier plus a cheap bulk tier.
  • Output tokens drive the bill. In the all-Astra row, 8M output tokens cost more ($320) than 100M input tokens ($268). Reasoning-heavy models that think longer cost more even at the same list price.
  • Gemini 3.8 Flash’s introductory price ends December 31, 2026. At $1.50 / $7.50 from January 1, the budget split’s implementer line becomes $69.30 and the total $80.24.

To run these numbers with your own token volumes and cache rate, use the AI pricing calculator.

The Router Pattern: Try Cheap, Escalate Failures

Splitting by role is one lever. The other is routing within a role: send each request to the cheapest tier first, check the result automatically, and only escalate what fails.

  1. Start on the cheap tier (GPT-6 Luna, Haiku 5.5, GLM-5.3 Flash, DeepSeek V4.1 Flash).
  2. Check the result with something that doesn’t need a human: tests pass, the JSON validates against your schema, the diff applies, the classifier’s confidence is above a threshold.
  3. Escalate only the failures to the mid tier, and the mid tier’s failures to the flagship.

The cost per finished task is the cheap attempt plus the failure rate times the next tier’s attempt. Example: a task with 40K uncached input tokens and 4K output tokens costs $0.006 on GPT-6 Luna (0.04M × $0.10 + 0.004M × $0.50) and $0.12 on GPT-6.2 Sol (0.04M × $2 + 0.004M × $10). If Luna fails 30% of the time and those go to Sol, the average is $0.006 + 0.3 × $0.12 = $0.042, about a third of running everything on Sol. The catch is the failures you don’t detect. Routing only saves money when the check in step 2 is reliable.

The same logic applies when comparing models directly: two models with the same list price can differ a lot in tokens per task. Measure cost per completed task on a sample of your own work before committing.

Caching and Batch Discounts That Apply

  • OpenAI: cache reads are $0.40 (Astra), $0.10 (Sol) and $0.01 (Luna) per 1M. 30-minute cache writes cost $10 / $2.50 for Astra / Sol. Batch costs 50% of Standard; Fast mode is 2x.
  • Anthropic: Opus 5.5 cache reads are $0.20, five-minute writes $5 (one-hour $8), Batch $2 / $10. Sonnet 5.5 cache reads are $0.10, writes $2.50 (one-hour $4), Batch $1 / $5. Haiku 5.5 Batch is 50% off.
  • Google: Gemini 4 Neon and Hydrogen use implicit caching only ($0.15 and $0.075 cached input). Batch / Flex: Neon $1.50 / $7.50, Hydrogen $0.75 / $3.75. Gemini 3.8 Flash Batch and Flex are half the synchronous rate.
  • DeepSeek: off-peak hours cost 50% less on both V4.1 Flash and V4 Pro, including cache hits.
  • Moonshot (Kimi K3.1 Pro): cache hits $0.20; cache writes are billed by TTL.
  • Alibaba (Qwen4 Max): implicit and explicit cache hits $0.075; explicit writes $1.875 (5-minute TTL) or $3 (1-hour).
  • Mistral: cached input tokens are 90% off.
  • Z.AI (GLM-5.3): cached-input storage is free for a limited time.

Batch fits work that can wait (nightly code review, bulk labeling, test generation). Caching fits agent loops that resend the same system prompt and repo context every turn, which is most of them.

Sources

Next step

Related Posts