Migrating to DeepSeek V4.1 Flash 1017: The reasoning_content 400, the Silent Alias Swap, and Peak-Hour Billing

Most migrations start with changing a model name. DeepSeek V4.1 Flash 1017 did it the other way round: on October 17, 2026, DeepSeek kept the name deepseek-flash and changed the model behind it. Anyone calling deepseek-flash, deepseek-v4-flash, or deepseek-v4-flash-vision-exp on DeepSeek's API was moved over without touching a line of code.

On AIHubMix the new checkpoint has its own ID, deepseek-v4.1-flash-1017, so you choose when to move. Either way, these are the problems to check:

AIHubMix ran a short set of test requests against deepseek-v4.1-flash-1017 on October 19, 2026 (03:02 UTC). Results from that run are marked as such below.

1. The 400 that breaks tool loops

What you'll see. The second request after a tool call fails with HTTP 400 and this message, as reported in a GitHub issue against opencodex for V4.1 Flash on October 15:

[invalid_request_error] The `reasoning_content` in the thinking mode must be passed back to the API.

The same text appears in earlier reports against opencode, LangChain, and claude-code-router for V4 Pro and V4 Flash.

Why. DeepSeek's thinking mode guide says that when a request includes tools, the reasoning_content of every earlier assistant turn must be sent back, "even for turns where the model did not perform a tool call". If it isn't, "the API will return a 400 error". Without tools, sending it back is optional and ignored. Clients usually trip on this because their message serialiser keeps content and tool_calls but drops the extra field.

The request that triggers it. Turn two of a tool loop, with the assistant message stripped of reasoning_content:

curl -s https://aihubmix.com/v1/chat/completions \
  -H "Authorization: Bearer $AIHUBMIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017",
    "tools": [{"type": "function", "function": {"name": "get_weather",
      "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}}}],
    "messages": [
      {"role": "user", "content": "What is the temperature in Hangzhou? Use the tool."},
      {"role": "assistant", "content": null, "tool_calls": [{"id": "call_00_example", "type": "function",
        "function": {"name": "get_weather", "arguments": "{\"city\": \"Hangzhou\"}"}}]},
      {"role": "tool", "tool_call_id": "call_00_example", "content": "{\"celsius\": 18}"}
    ]
  }'

On DeepSeek's own API, this is the shape that returns the 400 in the public reports. Through AIHubMix, the same shape returned HTTP 200 and a normal answer in the October 19 test run. That was one run, so don't build on it as a guarantee.

How to fix it. Keep the field, whichever route you use:

msg = resp.choices[0].message
messages.append(msg)  # carries content, reasoning_content, and tool_calls over

• In your own code, append the whole assistant message, not a hand-built copy of it.
• In frameworks, upgrade: opencode fixed the round trip in later releases, and opencodex merged its fix on October 16.
• As a last resort, set {"thinking": {"type": "disabled"}} for tool loops. That avoids the error but gives up the reasoning that makes the model good at agent work.

2. The same name, a different model

DeepSeek's changelog for October 17 says it plainly: "The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model." There is no pinned name for the September checkpoint, and the two older Flash names now run on 1017 too.

Nothing errors, but three things can shift:

• Output length. 1017 was trained to produce shorter outputs. Parsers, UI layouts, and prompts that expected the September model's long explanations may get less text than before.
• Evaluation baselines. Scores and costs you recorded on the September checkpoint no longer describe what deepseek-flash returns.
• Behaviour. DeepSeek retrained 1017 for end-to-end project work. Agents may take different paths to the same result.

What to do. Record the model and date with every evaluation run, and rerun your regression set now. The response tells you which checkpoint answered: in the October 19 test, every response through AIHubMix carried "model": "deepseek-v4-1-flash-261017". Log that field, and an unannounced swap becomes visible the day it happens.

3. Billing rules that are easy to miss

Peak and off-peak. DeepSeek charges full price 01:00–04:00 and 06:00–10:00 UTC on weekdays, and half price at every other hour, all weekend, and on Chinese public holidays. AIHubMix follows the same windows and prices each request at the time it was submitted, so a long streaming request that starts off-peak stays off-peak.

AIHubMix list prices per 1M tokens for deepseek-v4.1-flash-1017.

Cache hits are 50 times cheaper than misses. Agents resend the same system prompt, tool definitions, and history on every call. Keep that prefix stable, put anything that changes per call at the end, and most of each request bills at the cache-hit rate.

Thinking is billed as output. Thinking is on by default, and reasoning tokens count toward both the bill and max_tokens.

Search and Fetch. DeepSeek opened its new Search and Fetch APIs on October 17 at $0.75 per 1,000 successful calls each. Server-side calls made through DeepSeek Harness, Claude Code, or Codex on DeepSeek's API are free until October 31, 2026, 16:00 UTC. After that, an agent that searches on every turn adds a per-call cost on top of tokens.

What it adds up to. Under stated assumptions (40 calls, 1.2M input tokens with 90% cache hits, 60K output tokens), one agent session costs $0.118 at peak and $0.059 off-peak at AIHubMix list prices. Moving batch runs off-peak is the single biggest saving.

4. Settings that cost more than they return

Effort. The API offers low, high (default), and max. DeepSeek's V4.1 Flash paper maps them to internal effort values of 50, 75, and 100, and reports that the step to max "lengthens agent trajectories by 1.6–1.8× for only marginal improvements". DeepSeek's own Claude Code setup guide still sets CLAUDE_CODE_EFFORT_LEVEL=max. For everyday agent work, high is the better default.

DeepSeek also maps other effort names onto its three levels: minimal becomes low, while medium and xhigh both become high. A tool that offers five levels effectively gives you two or three. In the October 19 test, AIHubMix accepted low, high, and max on Chat Completions.

max_tokens. The model card recommends at least 256K for long agent runs, because thinking shares the output budget. With a small cap, a hard step can spend the whole budget thinking. Check for finish_reason: "length".

Sampling parameters. In thinking mode, temperature, presence_penalty, and frequency_penalty are accepted without error and have no effect. top_p is clamped to 0.95–1.0. If a pipeline relied on low temperature for deterministic output, it no longer gets it.

5. Agents with real permissions

1017 was trained with reinforcement learning on whole projects, so it tends to keep building until the job looks done. Its answers are also shorter, which means less narration of what it changed. Two habits help:

• Review diffs and command logs, not the model's summary of them.
• Run it in a sandbox with an allowlist of commands, and require approval for anything destructive (deletes, force pushes, migrations, outbound network calls).

DeepSeek itself flags the Flash line's strength on security tasks; the September checkpoint scored 88.1 on CyberGym in DeepSeek's own table. That's useful for defensive work, and another reason to scope its permissions tightly.

6. Coding tools

AIHubMix serves deepseek-v4.1-flash-1017 on all three protocols the coding tools need. In the October 19 test, Chat Completions, Anthropic Messages (including a two-turn tool call), and the Responses API all returned 200.

Claude Code (Anthropic Messages). Point it at AIHubMix as in the Claude Code guide and set the model:

export ANTHROPIC_BASE_URL="https://aihubmix.com"
export ANTHROPIC_AUTH_TOKEN="$AIHUBMIX_API_KEY"
export ANTHROPIC_MODEL="deepseek-v4.1-flash-1017"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4.1-flash-1017"
export CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4.1-flash-1017"
export CLAUDE_CODE_EFFORT_LEVEL="high"

• Thinking: in the test, a Messages request with no effort setting came back without a thinking block, while one with output_config.effort returned thinking. Keep an effort level set so 1017 reasons, and check the request logs.
• Context: DeepSeek's guide adds a [1m] suffix to the model name for its own endpoint; that suffix was not tested through AIHubMix.
• Known issue: through some routers, Claude Code has dropped thinking content and hit the 400 above (claude-code-router issue 1378). The two-turn tool test through AIHubMix's Messages endpoint did not hit it.

Codex (Responses API). Follow the Codex CLI guide: an aihubmix provider with base_url = "https://aihubmix.com/v1" and wire_api = "responses", and model = "deepseek-v4.1-flash-1017". To pick it from /model, add it to a model_catalog_json file with context_window set to 1,000,000.

• Effort: the guide's catalog script offers low, medium, and high. Under DeepSeek's mapping, medium and high both land on DeepSeek's high, so you get two real levels.
• Known issue: newer Codex versions discard the whole catalog if a single required field is missing. The guide's script avoids that.

DeepSeek Harness (Chat Completions). DeepSeek says 1017 works best with Harness 0.2.1-rc.1 or later. Add AIHubMix as an OpenAI-compatible provider following the DeepSeek Harness guide and add deepseek-v4.1-flash-1017 as a model. That guide was written against an earlier Harness release, so check field names if the UI has changed.

A 15-minute smoke test. Create a tiny repo with one failing test:

mkdir flash-smoke && cd flash-smoke && git init -q
printf 'def mean(xs):\n    return sum(xs) / len(xs)\n' > stats.py
printf 'from stats import mean\n\ndef test_mean():\n    assert mean([1, 2, 3]) == 2\n\ndef test_empty():\n    assert mean([]) == 0\n' > test_stats.py

Start your tool in that folder and ask: "Run pytest, fix the failing test by changing stats.py only, then run pytest again." It passes if:

• the agent makes at least two tool calls (run the tests, edit the file) with no 400;
• both tests pass at the end;
• the AIHubMix request log shows deepseek-v4.1-flash-1017, and the session cost is a few cents at most.

Migration checklist

• Decide the model ID: deepseek-v4.1-flash-1017 on AIHubMix.
• Append whole assistant messages in tool loops so reasoning_content goes back every turn.
• Rerun your regression and evaluation set; record model, date, and the response model field.
• Check parsers and UIs for shorter outputs.
• Set effort to high by default; reserve max for the hardest tasks.
• Raise max_tokens for agent turns and alert on finish_reason length.
• Remove temperature-based determinism assumptions in thinking mode.
• Schedule batch and overnight agent runs off-peak.
• Keep prompt prefixes stable for cache hits.
• Budget Search and Fetch calls from October 31 if you use DeepSeek's server-side tools.
• Sandbox agents with real permissions and review diffs, not summaries.
• Run the 15-minute smoke test in each coding tool you use.

FAQ

Do I need to change code to use DeepSeek V4.1 Flash 1017?
On DeepSeek's API, no: deepseek-flash already points to 1017. On AIHubMix, change the model name to deepseek-v4.1-flash-1017.

Why do I get a 400 about reasoning_content?
DeepSeek requires the reasoning_content of every earlier assistant turn to be sent back when a request includes tools. Clients that drop that field fail on the call after the first tool call.

Does the reasoning_content 400 happen through AIHubMix?
It did not in AIHubMix's test run on October 19, 2026, where a turn without reasoning_content returned 200. Keep sending the field anyway, since it is DeepSeek's documented rule.

Can I still use the September V4.1 Flash on DeepSeek's API?
No. DeepSeek moved deepseek-flash and the older Flash names to 1017 and offers no pinned name for the September checkpoint.

When is DeepSeek V4.1 Flash cheaper?
Outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Every other hour, weekends, and Chinese public holidays cost half.

Which effort level should coding agents use?
High, the default. DeepSeek's own paper says max lengthens agent runs by 1.6 to 1.8 times for small gains.

Does 1017 work in Claude Code and Codex through AIHubMix?
Yes. AIHubMix serves it on the Anthropic Messages and Responses protocols those tools use, and both passed a basic test on October 19, 2026.

Keep reading: the DeepSeek V4.1 Flash 1017 series

• Not sure the move is worth it yet? The upgrade post shows what 1017 gains over the September checkpoint and what a session costs: DeepSeek V4.1 Flash 1017: Same Price, Shorter Outputs, and AIHubMix vs OpenRouter Pricing
• Setting up Claude Code anyway? See how 1017 compares with the model Claude Code uses for its own subagents: DeepSeek V4.1 Flash 1017 vs Claude Haiku 5.5: Which Cheap Model for Terminal Coding Agents?
• Migrating from V4 Pro rather than the old Flash? This post checks whether you still need Pro at all: DeepSeek V4.1 Flash 1017 vs V4 Pro: Is the Flash Enough Now?

Sources

• Thinking Mode (DeepSeek)
• Change Log (DeepSeek)
• Integrate with AI Tools (DeepSeek)
• DeepSeek V4.1 Flash 1017 on AIHubMix (AIHubMix)
• DeepSeek V4.1 Flash returns reasoning_content 400 after tool calls, issue 7118 (GitHub)