OpenAI DevDay 2026：開発者が注目すべき AI API 新機能 10

How to read this: shipped vs likely

Separate facts from keynote fiction so this is not a changelog rewrite. The OpenAI API changelog through late August 2026 is explicit: Assistants is gone; whisper-1 and related transcription models shut down 26 February 2027; Agent Builder, the old Evals product, and reusable prompt objects entered deprecation in June; GPT-5.6 Sol / Terra / Luna shipped; Fast mode replaced Priority; Ultrafast is still a limited preview; mTLS and X.509 workload identity went GA on 29 August; per-request regional processing, the Prompt Caching dashboard, spend limits, the Safety dashboard, and a Terraform provider are already in the console. None of that is a prediction.

DevDay is a consolidation day. Preview to GA, Beta disclaimer removed, enterprise features dropped to ordinary projects, several endpoints folded into one Responses built-in tool—that is the story a keynote likes. The useful question is not “will GPT-5.7 appear” but: which request surface becomes the only recommended path? Which JSON Schema flips from model hint to platform constraint? Which session state moves from your database into official Conversations?

This article scores APIs and developer tools, not consumer SKUs. Each prediction has evidence, an if-announced action, and a JSON layer. Odds are judgment, not a leak. If the room goes somewhere else, use the scorecard—you do not have to redo schemas you already split.

注目すべき API 方向 10

Sorted by how hard they hit an existing codebase, not by stagecraft. The first three force integration-surface changes; the middle four change cost, identity, and telemetry; the last three decide how voice and multimodal join the same JSON pipeline.

1. A dated Chat Completions sunset, Responses as the only first-class surface

Assistants is already dead. Chat Completions is alive, but 2026 features almost never land there: remote MCP, tool search, computer use, Skills, hosted shell, WebSocket Responses, and phase (commentary / final_answer) all hang off Responses. Reusable prompts never shipped on Completions. That is platform policy, not taste: the Agent-capable surface is being collapsed to one.

The likeliest DevDay line is not “one more Responses parameter” but a Completions freeze or shutdown date plus a migration guide in the same spirit as Assistants → Responses. Stop opening new Completions wrappers: tools, Structured Output, and session state should follow Responses input / output items. JSON takeaway: do not keep two tools[] shapes.

2. Self-serve hosted MCP and connectors; private tunnels for ordinary projects

GPT-5.5 already speaks MCP on Responses; the console has OpenAI-maintained connectors (Google, Dropbox, and more); the 19 May Secure MCP Tunnel lets ChatGPT, Codex, Responses, and AgentKit reach on-prem servers via a customer tunnel-client—mostly as an enterprise motion. The gap: typical projects still cannot one-click a private MCP, and the connector catalog does not cover homegrown tools.

Prediction: DevDay productizes hosted MCP / connectors and turns Tunnel into a project toggle. A nod to AAIF A2A interop is plausible—OpenAI already consumes MCP at the model layer, while sideways agent delegation is still a hole (see A2A vs MCP). Prepare Tool.inputSchema that shares origin with tools[].parameters. The platform will not runtime-validate your server. The chain stays parse → schema → business rules.

3. Programmatic tool calling and multi-agent orchestration leave Beta

On 9 July GPT-5.6 added programmatic tool calling, explicit prompt cache controls, persisted reasoning, max reasoning effort, Pro mode, and multi-agent orchestration (Beta) on Responses. The model can chain tools structurally and treat child agents as first-class objects. Beta’s usual fate is a DevDay GA with limits and an SLA.

If it graduates, your orchestrator must decide which hops the platform owns versus which stay in your loop. A new JSON family appears: intermediate child-agent artifacts. Do not dump them into the same file as final Structured Output or MCP arguments. Stamp a correlationId on every tool log now so GA-day reconciliation is possible.

4. Structured Output 2.0: streaming partial JSON and schema alignment with tools

Today’s Structured Output can pin a final object and still stream a broken prefix; tool parameters and response json_schema remain two files; $ref, dynamic enums, and large schemas still hit silent limits. Our AI JSON error guide said it already: locking shape is not locking meaning. That is a developer pain a keynote loves.

Predicted trio: (1) streaming Structured Output whose deltas are legal partial objects or JSON Patch, not truncated strings; (2) an official way to declare tool schema and output schema as the same origin; (3) a less lossy JSON Schema subset (fewer silent drops of oneOf / $ref). What you can do now: keep schemas/tools and schemas/output apart, ship strict + additionalProperties:false, and check platform errors against local errors in JSONVue.

5. Conversations as the unified session primitive across Responses and Realtime

Assistants threads are gone; Conversations is the official replacement. Realtime still has its own session; Responses can be a stateless call or a session hang. Three memories at once is the messiest 2026 integration layer: your DB, Conversations, and the Realtime session each hold user state.

DevDay may announce “one Conversation, many modalities”—text Responses, voice Realtime, even image tools on one timeline. Thread migration cannot slip; the item-list JSON is a contract, not debug residue. Get create / append / export working, then Diff the platform session JSON against your business state.

6. Ultrafast and Fast become generally available service tiers with an SLA

On 30 July Fast replaced Priority (the priority tag maps automatically). On 13 August Ultrafast entered limited preview for GPT-5.6 Sol—up to about 14× Standard, per OpenAI. On 5 August Fast started accepting long-context prompts over 272K tokens. Preview-to-GA is stock DevDay plot.

After GA, service_tier sits next to model as a routing decision: Ultrafast for latency, Batch for bulk, Pro / Standard for long reasoning. Expect per-tier limits and documented error codes. Put service_tier and prompt_cache_retention in the request JSON now, not in SDK defaults, or the day after the keynote you will fail to match invoices and cache hit rate.

7. Enterprise identity as the production default: mTLS, X.509, WIF, regional processing

Workload identity federation shipped 26 May; per-request regional processing arrived 21 August (Global geography projects); mTLS and X.509 federation went GA 29 August. Add project IP allowlists, spend limits, and usage-by-API-key, and the enterprise stack is already there. What is missing is a recommended default architecture, not a new protocol.

Keynotes love a single slide: short-lived tokens + certificate identity + regional host prefixes + hard spend caps. Action: stop treating a long-lived API key in env as the only identity; make region a per-request knob. Those metadata fields show up in usage JSON and audit logs—name them before the console screenshot becomes the spec.

8. Production observability succeeds old Evals: cache, safety, and traces as one product

The old Evals product entered deprecation on 3 June. Meanwhile the Safety dashboard (blocked requests by safety_identifier) shipped 23 June, Prompt Caching 20 August, and usage filtered by API key 4 August. Dashboards are filling the hole a evals suite left.

Prediction: DevDay bundles cache hits, safety blocks, and traces (plus a possible lightweight eval hook) as Production Observability. Send a stable safety_identifier on every Responses call and record cache-read / cache-write tokens yourself. Do not log only output_text; keep usage, cache, and block reason in one structured record so a new dashboard maps onto something you already store.

9. Realtime agents gain the Responses tool stack: MCP and Structured Output on voice

Realtime is not a beta leftover: Realtime-2 in May, Realtime-2.1 (and mini) in July, plus Translate and Live Transcribe; whisper-1 dies in February 2027. The gap is tool depth—MCP, tool search, and json_schema outputs that exist on Responses often need a hand-rolled bridge in a voice session.

The demo that plays on stage is a voice agent calling MCP and locking the final answer to the same JSON Schema. If that ships, voice and text must not keep two tool definitions. Share schema files now; run Realtime function-call arguments through the same validator as Responses. A thinner bridge means less work on announcement day.

10. Multimodal generation folds into Responses built-in tools with one result JSON

Image 2, Sora 2 (edits / extensions / Batch), and GPT Transcribe are still dedicated endpoints; Responses already hosts image generation, web search (including image results), and computer use. The platform wants “model + built-in tools,” not seven /v1/* URLs in your head.

Prediction: more video, transcription, and image-edit work appears as tools on the Responses timeline, with typed output items instead of per-endpoint top-level JSON. Define a media-result envelope (id, type, url, usage) in your domain layer so both dedicated APIs and Responses tools map into it. When shapes diverge, Diff the adapter—not the model’s mood.

A Responses skeleton that is valid today and should remain valid after DevDay—MCP plus Structured Output in one call:

{
  "model": "gpt-5.6-sol",
  "input": "Extract the invoice fields as JSON",
  "tools": [
    {
      "type": "mcp",
      "server_label": "finance",
      "server_url": "https://mcp.example.com"
    }
  ],
  "text": {
    "format": {
      "type": "json_schema",
      "name": "invoice",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "invoiceId": { "type": "string" },
          "total": { "type": "number" },
          "currency": { "type": "string", "enum": ["USD", "CNY", "EUR"] }
        },
        "required": ["invoiceId", "total", "currency"],
        "additionalProperties": false
      }
    }
  }
}

Make service tier, cache retention, and safety identifiers explicit fields now:

{
  "model": "gpt-5.6-sol",
  "service_tier": "ultrafast",
  "prompt_cache_retention": "24h",
  "safety_identifier": "user-42"
}

対照表：根拠、確度、用意する JSON

Ten rows, one decision table. Odds are the author’s: high = the changelog already points at consolidation; medium = preview/Beta exists but productization is unclear; low = directionally right, maybe only a breakout slide.

Low-odds side bets that still matter: wider third-party model routing (Bedrock already offers an OpenAI-compatible Responses endpoint), deeper Codex / Agents SDK ties to the console, and read-only A2A Agent Card compatibility. They will not eat twenty keynote minutes; a breakout could still redraw your integration map.

DevDay 前 2 週間でできること

Do not rewrite the product for a forecast. Do four things that pay off even if every prediction misses:

1. Freeze a canonical schema layout: tools, output, and session items in three directories—no single giant JSON.
2. New code on Responses only; mark Completions wrappers stale. Clear leftover Assistants objects using the migration guide.
3. Put reconcilable fields on every request: safety_identifier, your request_id, a planned service_tier.
4. Keep one real failure fixture (truncated JSON, schema error, drifted MCP arguments) so announcement-day diffs have a baseline.

You can run the loop in the browser:JSON 整形for parse;JSON Schema 検証for tools and output;JSON Difffor model arguments vs MCP params.arguments. Nothing leaves the machine.

Further reading: Assistants → Responses migration, Structured Output, AI JSON errors, MCP and JSON Schema, A2A vs MCP.

FAQ

まとめと次の一歩

What matters at DevDay 2026 is not another model name. It is whether the platform collapses Responses, MCP, sessions, tiers, and identity into a default stack. All ten predictions say the same thing: fewer parallel APIs, more JSON you actually have to honor.

Before 29 September, stop new Completions code, split the schemas, keep a failure fixture. On the day, read the changelog—not a social recap. When you need to check a payload shape, open JSONVue.