Frontier coding models in October 2026: prices, limits and how to choose

The strongest coding models of October 2026 differ in price by more than 20x on output, and their context windows, vision support and API quirks differ too. This guide sets out the facts from each maker's own pages, then makes a practical case: use a cheaper model by default and escalate to a frontier coding model only when the task needs it.

LADDER = [
  "deepseek/deepseek-v4.1-flash", # default
  "anthropic/claude-sonnet-5.5",  # on fail
]
# check exact ids: GET /v1/models

The frontier coding models at a glance#

List prices checked 3 October 2026, in USD per million tokens, from each maker's site. "Positioning" is the maker's own description, quoted or closely paraphrased.

* GPT-5.6 Sol's price is promotional. The launch price was $5 / $30. OpenAI cut it on 21 August 2026 and guarantees the lower price at least until 21 November 2026. It applies to prompts up to 272K tokens. We didn't confirm image input for Sol in our check.

Notes model by model#

Claude Sonnet 5.5#

Released on 28 September 2026. Anthropic charges the standard rate across the full 1M context, with no surcharge for long prompts. Prompt caching costs $0.20 per million for reads, and $2.50 (5-minute) or $4 (1-hour) for writes. The Batch API halves the price to $1 / $5. Adaptive thinking is on by default at high effort.

Two API details catch agent frameworks out:

• Forced tool choice (any or a named tool) isn't supported. Use auto.
• Setting temperature, top_p or top_k to a non-default value returns an error. If your agent sets a temperature, remove it for this model.

GPT-5.6 Sol and the GPT-6 generation#

OpenAI names its tiers Sol (flagship), Terra (balanced) and Luna (cheap, high-volume). GPT-5.6 Sol was released on 9 July 2026. OpenAI already lists successors at lower prices: GPT-6 Sol at $2 / $10, and GPT-6.1 Sol at $2 / $0.10 cached / $10. Those are half of GPT-5.6 Sol's promotional price.

The cheap model in the new generation, GPT-6 Luna, has the same 1,050,000-token window. It costs $0.10 / $0.50 up to 272K input tokens and $0.20 / $0.75 above that.

If you're choosing an OpenAI flagship today, check which of these your provider actually serves before you settle on GPT-5.6 Sol. Our catalog is at /models.

Grok 4.7#

Released on 21 September 2026. It has the smallest window in this group, at 500K. The price doubles to $4 / $12 for prompts of 200K tokens or more. It takes text and image input and supports function calling and structured outputs.

Kimi K3#

Moonshot's flagship. Thinking is always on, with effort levels low, high and max, and max is the default. That's the most expensive setting on a model that already charges $15 per million output tokens, so for routine turns lowering the effort is the first thing to try.

Caching has a separate write charge: $3 per million for a 5-minute TTL and $6 for 1 hour, on top of $0.30 per million for cache reads. The output limit defaults to 131,072 and can be raised to the full 1,048,576. Moonshot doesn't tier prices by context length.

GLM-5.3#

Z.ai's flagship for coding and agents, released on 18 August 2026. It's text-only and reasoning is always on. GLM-5.1, 5.2 and 5.3 are all priced at $1.40 / $4.40, so unless you need GLM-5.2's option to turn thinking off, there's no price reason to use an older version.

Qwen 3.8 Max#

The qwen3.8-max alias has pointed at the qwen3.8-max-0902 snapshot since 5 September 2026. Alibaba says the snapshot improves coding and multi-tool use, at the same price.

Cache pricing has three rates:

• Implicit cache hit: $0.25 per million.
• Explicit cache read: $0.17.
• Explicit cache creation: $2.50.

The input limit is 991,808 tokens (983,616 in thinking mode), slightly under the "1M" headline. The price is flat; it doesn't change with prompt size.

DeepSeek V4 Pro#

DeepSeek's flagship. It's text-only, with thinking effort levels low, high and max, and it supports the Responses API natively. DeepSeek uses the same peak and off-peak hours as for V4.1 Flash. In Dhaka time, peak is 07:00 to 10:00 and 12:00 to 16:00 on weekdays.

There's an availability risk. In September, DeepSeek announced that V4 Pro requests would be routed to V4.1 Flash. It later changed its changelog to say V4 Pro would keep running "in response to user demand". Use it, but don't build anything that can't fall back.

MiMo V2.6 Pro and MiniMax M3#

These are the two cheapest models in the table. MiMo V2.6 Pro (released 22 September 2026, MIT weights) charges $0.87 per million output tokens and $0.0036 for cache hits, and accepts text, image, video and audio. Xiaomi also sells a "UltraSpeed" version it says runs "up to 20x faster", at ten times the price.

MiniMax M3 doubles its price above 512K input tokens and only guarantees 512K of context. It sits in both this guide and the cheap-models guide, which tells you something about its pricing.

How to choose a frontier coding model#

Rule out models before you compare them:

1. Do you send screenshots or diagrams? GLM-5.3 and DeepSeek V4 Pro are text-only.
2. Do your sessions go past 500K tokens? Grok 4.7 can't take them. Even below that, check the tier prices in the million-token context guide.
3. Does your framework force tool choice or set temperature? Claude Sonnet 5.5 rejects both.
4. Do you need a model you can rely on for months? DeepSeek V4 Pro's retirement was announced once already, and GPT-5.6 Sol's price is only guaranteed until late November.

Among what's left, compare cost per completed task rather than price per token. A model that's three times cheaper per token but needs four attempts costs you more. You only find that out by running your own tasks and reading the usage numbers. We aren't going to rank these models for you without that data.

Start with a cheap model, escalate to a frontier one#

Many agent turns are routine. A pattern that keeps costs predictable is to try a cheaper model first, check the result with something objective like your test suite, and only send the task to a frontier model if the check fails. Here's a minimal version against the OpenAI-compatible endpoint:

import os, subprocess
from openai import OpenAI

client = OpenAI(base_url="https://tokens.bd/v1",
                api_key=os.environ["TOKENS_API_KEY"])

LADDER = ["deepseek/deepseek-v4.1-flash",   # check ids in /models
          "anthropic/claude-sonnet-5.5"]

def solve(task, apply_patch, revert):
    notes = ""
    for model in LADDER:
        resp = client.chat.completions.create(
            model=model,
            messages=[{"role": "user", "content": task + notes}],
        )
        apply_patch(resp.choices[0].message.content)
        tests = subprocess.run(["npm", "test"], capture_output=True, text=True)
        if tests.returncode == 0:
            return model
        revert()
        notes = "\n\nA previous attempt failed these tests:\n" + tests.stdout[-4000:]
    return None

Some practical points:

• Pass the failure on. The escalated request includes the test output, so the stronger model starts from what went wrong, not from scratch.
• Don't set temperature in a ladder that includes Claude Sonnet 5.5.
• Switching models loses the prompt cache. Caches are kept per model, so the escalated request pays full input price for the whole context. That's one more reason to escalate once per task, not once per turn.
• Restrict the key. Give the key an allowed-models list containing just the two ladder models and a monthly spend cap. Both are set when you create the key; see API keys. A request for any other model gets 403 model_not_allowed_on_key.

The exact model IDs on Tokens are in provider/model form. Check them in the catalog or with GET /v1/models, which also shows only the models your plan can call. Choosing a model covers the same decision from the configuration side.

List prices checked 3 October 2026. Promotional prices and model status change often; the maker's page is the authority.

Sources: Anthropic pricing · Claude Sonnet 5.5 overview · OpenAI pricing · xAI models · Kimi pricing · Kimi models · Z.ai pricing · GLM-5.3 · Alibaba Model Studio pricing · Qwen 3.8 Max · DeepSeek pricing · DeepSeek updates · Xiaomi MiMo · MiniMax pricing