Model benchmarks, use cases and setup: November 2026

Go to: CrucibleForge on GitHub

Which language model should you use in November 2026, for which job, and how do you set it up? This guide answers that from eleven models scored on the same tests in MakoBench 4.1. The full board ranks more models; these eleven are the ones we compared in depth this month. You get the scores, the pick for coding and for chat, a five-step setup for a lead-plus-coder stack, and what one run or a whole project costs.

In one screen

• Best overall: a five-way tie at the top. Kimi K3.1 Pro, Claude Fable 5.2, Grok 4.7, GPT-6 Astra and GPT-6.1 Sol finish between 23.9 and 25.2. Claude Fable 5.2 is listed first, but the five are in a statistical tie.
• Best for agentic coding: Claude Fable 5.2 when money is no object: the top Work score (12.5), a perfect agentic pack, and 100% on the standard agent cases. GPT-6 Astra or GPT-6.1 Sol for less: both have a perfect agentic pack, but 66.7% on the standard agent cases.
• Best cheap coder: DeepSeek V4.1-Flash-1017. Standard coding 10.0, $0.14 a run.
• Best for a chat or help line: GPT-6.1 Terra on an API (chat 9.2, $0.63 a run), or Swish 31B Final on your own hardware. The best chat score outright is Claude Fable 5.2's 12.7, at $14.87 a run.
• Best on your own hardware: Swish 31B Final, 16.3 overall, no API bill, for chat or anything else that must stay local.
• Recommended stack: Claude Fable 5.2 plans and reviews, DeepSeek V4.1-Flash-1017 writes the code. Estimates are $4.82 for a ChatForge-sized build and $57.86 for a DisPatch-sized one.
• Corrected: Claude Fable 5.2's three agentic tests first scored 0 because our API account ran out of credit mid-run. Re-run, it passed all three, and its score went from 24.2 to 25.2.

Pick a model in two questions

Start with the job, then ask one question. For code, ask about budget: a big budget means Claude Fable 5.2, and a tight budget means DeepSeek V4.1-Flash-1017, $0.14 a run. For chat, ask whether it has to run on your own hardware: on an API, GPT-6.1 Terra, $0.63 a run; on your own hardware, Swish 31B Final. The picks are the same ones the rest of this guide backs up with numbers.

What is the job?

↓

Code

Budget?

big budget

↓

Claude Fable 5.2most expensive

tight budget

↓

DeepSeekV4.1-Flash-1017

Chat

Own hardware?

no

↓

GPT-6.1 Terracheap chat

yes

↓

Swish 31B Finalno API bill

The November 2026 board

Correction: an earlier version of this guide gave Claude Fable 5.2 24.2 and called it provisional. Its three agentic tests had scored 0 because our API account ran out of credit mid-run, not because the model failed them. We re-ran them through Anthropic's own API with prompt caching, for $10.81. It passed all three, which lifted its Work from 11.5 to 12.5 and its Total from 24.2 to 25.2.

Total is Work plus Chat. Work is the board's Code column. The number includes the bonus packs a model unlocked, so it is not a score out of 20. The standard tests cap Work and Chat at 10 each, so the standard total caps at 20 and the published Total runs up to 28. Adult is a separate test and is not part of Total. A higher number there means the model went further on adult fiction between adults, not that it is a better model. Code bank / Agent is those two shares of the coding tests, not the programs score. Run cost is one standard Code and Chat run at the provider's price, without the packs, the adult test, or retries. "local" means our own hardware, no API bill. Maxed packs are packs scored at a perfect 1.000. No unlocked pack in this set is waiting to be scored. Rows that ran out or errored still count as failed, and the rank stands: DeepSeek V4-Pro-0813 has 11, DeepSeek V4.1-Flash-1017 has 5, MiniMax M3.1 Flash has 7 plus one budget stop, Kimi K3.1 Pro has 2, Claude Fable 5.2 has 1, and Grok 4.7 had 1 row error.

Bonus packsadd to Work and Chat

↓

Workthe Code column

Chatthe chat tests

↓

TotalWork plus Chat

Not part of Total

Adult testa separate score

The top of the board is a tie, not a ranking. Claude Fable 5.2, Kimi K3.1 Pro, Grok 4.7, GPT-6 Astra and GPT-6.1 Sol finish between 23.9 and 25.2. The board lists them 1 to 5, with Claude Fable 5.2 first, but those gaps sit inside its ± margins, and its own rule calls models that close a statistical tie. Table order is not a winner. The next group, GPT-6.1 Terra, DeepSeek V4.1-Flash-1017 and DeepSeek V4-Pro-0813, sits between 18.4 and 19.4. Swish 31B Final, the top local, is 16.3. Coding on the standard test is already saturated for the leaders: Kimi K3.1 Pro, Claude Fable 5.2, Grok 4.7 and DeepSeek V4.1-Flash-1017 all sit on the standard Work ceiling of 10.0, so the 24-to-25-point cluster is packs and chat, not a higher standard coding score. What separates that cluster from DeepSeek V4.1-Flash-1017 is chat, 12.4 to 12.7 against 7.6.

Three things stand out. Cost spans two orders of magnitude on the same work: one standard run costs $14.87 for Claude Fable 5.2, $5.70 for GPT-6 Astra, $3.91 for Grok 4.7 and $0.14 for DeepSeek V4.1-Flash-1017. The adult test splits the field the other way. Claude Fable 5.2 (12.0), Swish 31B Final (11.8), Grok Build 0.1 (11.7), Grok 4.7 (11.5) and DeepSeek V4.1-Flash-1017 (10.0) refused none of the nine adult cases. The three OpenAI models stop on adult fiction at 1.3 to 1.6. Kimi K3.1 Pro refuses four of nine and scores 0 on adult writing. Those are the models declining, not a provider filter: no reply in this set was blocked by a content filter. Third, the agent cases are the coding gap, and the speed spread is wide. All three OpenAI models score 66.7% on the agent cases, where Kimi K3.1 Pro, Claude Fable 5.2, Grok 4.7 and DeepSeek V4.1-Flash-1017 score 100%. GPT-6.1 Terra is the fastest model in the set. Grok Build 0.1 runs at 8.4 tokens per second, slower than Swish 31B Final at 41.3 on the same board.

Use cases: which model for which job

Agentic coding of new software

• Money is no object: Claude Fable 5.2. Pro: the best chat score in this set (12.7), the top Work score (12.5), listed first in the tie for the top Total, a perfect agentic pack, 100% on the standard coding percents, and no refusals. Con: $14.87 for one standard run, the most expensive model here.
• A perfect agentic pack for less: GPT-6 Astra, or GPT-6.1 Sol. Pro: like Claude Fable 5.2, both have a perfect agentic pack, at $5.70 and $1.17 a run. Con: both score 66.7% on the standard agent cases, and both stop on adult fiction.
• In the middle: Grok 4.7. Pro: Work 11.7, third in this set after Claude Fable 5.2's 12.5 and Kimi K3.1 Pro's 12.0, chat 12.4, and $3.91 a run, about a quarter of Claude Fable 5.2's $14.87. Con: metered at $2 per million input and $6 per million output, and it counted 696,485 reasoning tokens on the speed sample, so a long agent loop is not cheap.
• On the cheap: DeepSeek V4.1-Flash-1017. Pro: standard coding 10.0, the ceiling, and a pack-inclusive Work of 11.6, close to the leaders' 11.5 to 12.5, for $0.14 a run. Con: chat is 7.6, so anything user-facing needs a second model. The $0.14 figure is the board's bill at the older rate. That historical run stays at $0.14. A new job at the published list rate costs more.

Chat bot or customer-service line

A help line does not need a high adult score. Pick on chat, price, and whether the line has to stay on your hardware.

• GPT-6.1 Terra. Pro: chat 9.2 at $0.63 a run and $1 per million input, $5 per million output. That is the cheap chat-capable pick, and it is the fastest model in the set. Con: 66.7% on the agent cases. Programs and the code bank are both 100, so this is an agent miss, not a blank coding score. Its own speed varied widely across samples, from 72 to 1,044 tokens per second.
• Grok 4.7. Pro: chat 12.4, near the top of this set, at $3.91 a run. The two deep role-play packs are 0.945 and 0.951. Con: the conversation pack is 0.814 and the character pack is 0.483, so a persona that has to hold is the weak spot. Metered at $2 and $6 per million tokens. A 15x plan exists. The $3.91 is the metered list, not that plan.
• Claude Fable 5.2. Pro: the best chat in this set (12.7), the best single chat pack at 0.964, no refusals, and no reasoning tokens billed on this board. Con: $10 and $50 per million tokens, tied with GPT-6 Astra for the highest token price in the set.
• If the line has to stay on your own hardware: Swish 31B Final. Pro: 16.3 overall, no refusals, no API bill. Con: the agent test is 33.3%, Work is 8.4, and it runs at 41.3 tokens per second.

Setup: harness, lead agent, coder

For real work, use a harness, one strong model that plans and reviews, and a cheap model that writes the code. DeepSeek Harness (dsh) is the harness. Claude Fable 5.2 is the planner and the reviewer. DeepSeek V4.1-Flash-1017 is the coder: standard coding is already at the ceiling, and the board run is $0.14. MiniMax M3.1 Flash on its Token Plan can take bulk inside the window, with a Claude Fable 5.2 review on the output. Its chat of 5.6 is the weakest in this set, and its Work of 8.0 against 11.6 means more review loops. Grok 4.7 is the alternative when you want one model for everything. There are five steps.

dshruns the loop

↓

Lead: Claude Fable 5.2plans and reviews

↓

CoderDeepSeek V4.1-Flash-1017writes the code

Options

MiniMax M3.1 Flashbulk, inside its window

Grok 4.7one model for all

Step 1: Check the board and pick the ids

/makobench/check the board

↓

Four idsat the top of the script

↓

LEAD_MODEL

CODER_MODEL

BULK_MODEL

ONE_MODEL

The model ids sit at the top of the script. Change those, not a second copy further down. Check the board at /makobench/ before you do.

Step 2: Set the API keys

ANTHROPIC_API_KEYrequired

DEEPSEEK_API_KEYrequired

MINIMAX_API_KEYrequired

XAI_API_KEYoptional

The script needs ANTHROPIC_API_KEY, DEEPSEEK_API_KEY and MINIMAX_API_KEY. XAI_API_KEY is optional and only for the one-model option. Set them in your shell before you run it. The script stops if a required one is missing.

Step 3: Run the script

setup.shplus your project path

↓

models.ymlwritten to a temp dir

↓

No run startsit only writes the ids

The script writes the ids to a models.yml file in a temporary directory. It does not start a run. Pass it the path to your project.

#!/usr/bin/env bash
# Do-it-all coding setup: harness, lead agent, sub-agents.
# Check /makobench/ for the current leaders before changing any id below.
set -euo pipefail

# Board labels. dsh may show a different provider string for the same model.
LEAD_MODEL="claude-fable-5-2"        # Claude Fable 5.2, plans and reviews
CODER_MODEL="deepseek-flash"         # DeepSeek V4.1-Flash-1017, writes the code
BULK_MODEL="MiniMax-M3.1-Flash"      # MiniMax M3.1 Flash, window-limited bulk
ONE_MODEL="grok-4.7"                 # Grok 4.7, optional, one model for all

: "${ANTHROPIC_API_KEY:?set ANTHROPIC_API_KEY}"
: "${DEEPSEEK_API_KEY:?set DEEPSEEK_API_KEY}"
: "${MINIMAX_API_KEY:?set MINIMAX_API_KEY}"
: "${XAI_API_KEY:=}"                  # optional, only for ONE_MODEL

# Set this to the DeepSeek provider id your dsh lists. Do not guess it.
DSH_PROVIDER="${DSH_PROVIDER:?set DSH_PROVIDER}"

PROJECT="${1:?usage: setup.sh /path/to/project}"
JOB="$(mktemp -d)"

cat > "$JOB/models.yml" <<YAML
lead: ${LEAD_MODEL}
coder:
  provider: ${DSH_PROVIDER}
  model: ${CODER_MODEL}
bulk: ${BULK_MODEL}
one_model: ${ONE_MODEL}
YAML

echo "Lead: ${LEAD_MODEL}"
echo "Coder: ${CODER_MODEL} via ${DSH_PROVIDER}"
echo "Bulk, inside the plan window: ${BULK_MODEL}"
echo "One-model alternative: ${ONE_MODEL}"
echo "Id file: ${JOB}/models.yml"
echo "Project: ${PROJECT}"
# Point dsh at CODER_MODEL. Do not hardcode a second model id here.

Step 4: Point dsh at the coder

dshyour harness

↓

DSH_PROVIDERwhatever your dsh lists

↓

deepseek-flashthe coder id

Point dsh at the coder id. dsh's provider string is whatever your install lists for DeepSeek. The board's API id for the coder is deepseek-flash.

Step 5: Schedule around limits

DeepSeek peak, UTC

00:0012:0024:00

Monday to Friday

01:00 to 04:00

06:00 to 10:00

Off-peakhalf price

DeepSeek is half price off-peak. Peak is Monday to Friday 01:00 to 04:00 and 06:00 to 10:00 UTC. Schedule the heavy coding outside it.

MiniMax plan5-hour and weekly

↓

402the window is spent

↓

Switch the bulk modelDeepSeek V4.1-Flash-1017

The MiniMax plan is flat. Marginal API cost is $0 until a 5-hour or weekly window is used up. A 402 means the window is spent, not that the plan ended. Switch the bulk model to DeepSeek V4.1-Flash-1017 for the rest of the build. A 60,000-line build will spend 5-hour windows, so the plan is a throttle, not a free pass.

What a project costs

These are estimates from the board prices, not measurements from a real project. No token log for a project this size was available. The assumptions: finished source costs 15 tokens a line, so a ChatForge-sized project of 5,000 lines is 75,000 tokens and a DisPatch-sized project of 60,000 lines is 900,000 tokens. An agent re-reads its own source, so the mid band counts 12 times the source in and 3 times out, and a thinking coder uses the high band of 20 and 5. A fresh project is a cache miss. DeepSeek is priced at the live off-peak list rate. This is one build, not maintenance, and it excludes the judge and image generation.

Lines of code5,000 or 60,000

↓

Source tokens15 tokens a line

↓

Agent re-reads sourcemid band: 12 in, 3 outhigh band: 20 in, 5 out

↓

Price per millionboard prices

↓

Project costone build, an estimate

Claude Fable 5.2 spent no reasoning tokens on this board, so the mid band is the right figure for it. If the agent thinks all the way through, the high band is $33.75 and $405. DeepSeek V4.1-Flash-1017's high band is $0.45 and $5.40. Peak is double the off-peak rate, so the mid band at peak is $0.54 and $6.48. That is not the high band. In the mixed setup Claude Fable 5.2 takes 20% of the input and 25% of the output and DeepSeek V4.1-Flash-1017 does the rest, which lands at $4.82 and $57.86 off-peak. Using MiniMax M3.1 Flash as the coder instead drops the coder cost to $0 inside the plan window, but its Work of 8.0 against 11.6 means more review loops, and those loops are not in the $4.61 or the $55.35.

Recommendation

For something like DisPatch or ChatForge, use the mixed stack. Claude Fable 5.2 plans and reviews. DeepSeek V4.1-Flash-1017 writes the code. DeepSeek Harness runs the loop. GPT-6.1 Terra answers users. The mid estimate for that build is $4.82 for a ChatForge-sized project and $57.86 for a DisPatch-sized one, against $20.25 and $243 for Claude Fable 5.2 alone and $0.27 and $3.24 if DeepSeek V4.1-Flash-1017 does the whole job. The top of the board is tied, so pick the cheapest model that does the job, and keep the expensive model for planning and review. The $4.82 is the build, not the chat line. Terra alone, at the same assumptions, is $2.03 for a whole ChatForge-sized build and $24.30 for a DisPatch-sized one. Those are complete builds on Terra by itself, an alternative to the $4.82 stack: you pay one or the other, never both. Keep the MiniMax Token Plan for bulk work inside its window and treat it as a throttle. Do not pick DeepSeek V4-Pro-0813 over DeepSeek V4.1-Flash-1017. It scores lower on this board, 18.4 against 19.2, and costs more. If you would rather have one model than a stack, Grok 4.7 is the one: Total 24.1, chat 12.4, project estimates $3.15 and $37.80. Its board run is $3.91, about a quarter of Claude Fable 5.2's $14.87.

Maxed packs, and what comes next

Four models in this set have a pack scored at a perfect 1.000:

• Claude Fable 5.2: work/reasoning #1, work/agentic #1 and adult writing #2.
• Grok 4.7: work/reasoning #1 and adult writing #2.
• GPT-6 Astra: work/agentic #1.
• GPT-6.1 Sol: work/agentic #1.

Nobody scored 1.000 on the code pack. The best was GPT-6 Astra at 0.917. No chat pack reached 1.000 either. The best was Claude Fable 5.2's chat/deep_rp #1 at 0.964. The MAX mark on the board is a different thing. It says a part reached the top of its ladder, not that a pack was perfect. That is why GPT-6 Astra carries the code mark while its code pack share is 0.917. Acing a pack does not retire it. MakoBench 4.5 is now in progress and adds harder packs for full testing, so these 4.1 results are the current maximum, not the last word.

Every score, the test list and the run cost of each model are on the MakoBench board.