Open weights inside a closed harness

01The server connected; the tool was missing

A 31 July report on the Codex tracker describes a tool-visibility failure on Codex 0.145.0. After DeepSeek’s official setup, MCP handshakes succeeded and resources could be listed, yet the model had no callable MCP tools. The reported catalogue combined supports_search_tool: true with tool_mode: null: tools became deferred without a discovery tool being exposed. ⁶

Imagine asking Codex to inspect a database schema through an MCP server and propose a migration. The cheaper model replies fluently, reads some source files and suggests a patch. If it never called the schema tool, you have changed the task as well as the model. That is a hypothetical comparison, not a test we ran.

The integration examined in this 8 August edition made that a practical concern. DeepSeek supplied an official Codex setup that writes a model catalogue and provider configuration. The current guide describes backing up the configuration while preserving MCP servers and project trust settings. A working configuration gets requests to the provider; it does not prove that a particular tool reaches the model. ¹

Update, 18 October: the old claim that Codex could reach only V4 Flash is no longer current. DeepSeek’s guide now offers V4 Pro too. The August availability label in the preserved table below is historical. The argument here concerns the tool round trip, not a continuing inability to select Pro. ¹

02A listed server can still supply no callable tool

For our schema task, a successful server connection would therefore answer the wrong question. The useful evidence is the schema-tool invocation and its returned result in the session. A model saying it lacks the tool is also useful evidence; treating its substitute answer as equivalent hides the missing input.

The reporter found that disabling supports_search_tool exposed the tools in a new session. That is an attributed workaround for the documented configuration, not a blanket instruction for every current Codex installation. Preserve the client version and catalogue when reproducing the case. ⁶

03The request and the next turn have contracts too

A separate May report, #23186, identifies MCP tools wrapped in a namespace object on custom-provider Responses requests. Its examples distinguish backends that skip the wrapper, reject it, or return text without a function call. A provider accepting ordinary requests can still handle this tool schema differently. These are the reporter’s cases, not our cross-provider benchmark. ⁵

Issue #24500 reports a different symptom: a later thinking-mode request fails with HTTP 400 because reasoning_content is absent. The thread also questions where the conversion occurs. A commenter points out that DeepSeek’s cited replay requirement concerns Chat Completions history, so a Responses-to-Chat proxy may own the missing field. The report does not establish that every native Responses route has that fault. ³

In the schema example, tool visibility, request translation and the next turn are three separate observations. Seeing the first answer arrive proves little about the second request. A visible error is easier to identify than a missing tool, but neither result should count as a completed migration task.

04Keep the August price card in its own date

The table preserves the prices recorded for this edition on 9 August. DeepSeek’s maintained integration repository also retains those flat rates as a historical reference. OpenAI’s launch page records Sol at $5 input and $30 output, and notes a later August price reduction; Anthropic documents the Opus 5 rates. These are not today’s shopping recommendations. ^{11 9 10}

DeepSeek’s live pricing page has since changed model names and rates, and now lists Responses support for both models. Read that page for a purchase made today; the table’s old availability label belongs to the August case. Replacing the old figures silently would erase what the original comparison meant. ²

05The loop determines how much of the discount survives

Microsoft Research’s July account analyzes sampled June Copilot traces covering 3.2 million users and 13 million sessions. Model calls were coupled nearly one-to-one with tool execution. Cache hit rates averaged 90% within a turn and 55% across turn boundaries; model switches and compaction could invalidate them. This describes Copilot’s sampled workload, not a measured DeepSeek penalty. ⁴

Our hypothetical schema task could retry, find a different source or ask a human when the intended tool is missing. Those routes consume different work. The published trace study gives a reason to count calls and cache usage; it supplies no multiplier for how much this particular integration would cost.

A useful comparison would keep the repository, migration request and acceptance check the same. First require a real schema result and a completed tool round trip. Then compare total billed input, cached input, output, retries and review effort across repeated runs. Keep failed attempts in the comparison. We have not performed that experiment here.