feat: Add 6 new OpenRouter models (April 2026 refresh) by timeleft-- · Pull Request #430 · BeehiveInnovations/pal-mcp-server

timeleft-- · 2026-04-14T21:12:20Z

Summary

Add Nous Research Hermes 4 (405B & 70B) — hybrid reasoning models with deliberative alignment, function calling, and structured output
Add Qwen 3.6 Plus — Alibaba's latest frontier model with 1M context and agentic coding
Add MiniMax M2.7 — 1T parameter model matching GPT-5.3-Codex on SWE-Pro
Add Qwen3 Coder 480B — frontier open-weight coding model (35B active MoE)
Add Google Gemini 3.1 Flash Lite Preview — cost-efficient flash variant with thinking support
Migrate qwen alias from Qwen 3.5 → Qwen 3.6 Plus, minimax alias from M2.5 → M2.7

Total models: 35 → 41.

Test plan

JSON parses cleanly
Registry loads all 41 models with no alias conflicts
Unit tests pass (python -m pytest tests/ -v -m "not integration")
OpenRouter model simulator test (python communication_simulator_test.py --individual openrouter_models)

🤖 Generated with Claude Code

New models: Claude Opus 4.6, Claude Sonnet 4.6, Gemini 3.1 Pro, GPT-5.4 Pro, GPT-5.3 Codex, DeepSeek V3.2, Devstral 2512, Qwen 3.5 397B, MiniMax M2.5. Removed: meta-llama/llama-3-70b (8K context, obsolete), perplexity/llama-3-sonar-large-32k-online (legacy). Generic aliases (opus, sonnet, pro, gpt5pro) migrated to newest model versions. Version-specific aliases preserved for backward compatibility (opus4.5, sonnet4.5, gemini3.0, gpt5.2-pro). Version bumped to 9.9.0. All 870 tests pass. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Keep version 9.9.0 from our feature branch (upstream synced to 9.8.2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- gpt-5.3-codex: supports_temperature true→false (contradicts fixed constraint) - minimax-m2.5: max_output_tokens 196608→32768 (practical default; API allows up to 196K) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

10 files were failing the black --check CI step before this PR. None are files modified by this PR — fixing them here to unblock CI. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Add openai/gpt-5.4 with aliases gpt5, gpt5.4, gpt-5.4 - Move gpt5 alias from openai/gpt-5 to openai/gpt-5.4 (newest base) - Keep gpt-5.0/gpt5.0 on old openai/gpt-5 for backward compat - Update tests and docs Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- codex → openai/gpt-5.3-codex (was gpt-5-codex) - Old gpt-5-codex keeps codex-5.0 for backward compat Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Claude Desktop has a ~4 minute hardcoded MCP tool call timeout. When Pal routes to slow models (e.g. high-reasoning models), API calls can take 4-8 minutes and get cancelled before the response arrives. This wraps tool.execute() calls with periodic MCP progress notifications (every 15s) that keep the connection alive. Uses the standard MCP progressToken mechanism — no changes needed to individual tools or providers. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…26-03'

Address code review feedback from GPT 5.3 Codex: - Use `is None` instead of truthiness check for progress_token (0 is valid) - Wrap entire while loop in try/except CancelledError so cancellation during send_progress_notification also cleans up the task - Await the cancelled task to prevent orphaned coroutines - Log first notification failure instead of silently swallowing all Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Claude Desktop never sends a progressToken in MCP tool call requests, so progress notifications were never firing. Use request_id (always available) as fallback token. The notification still keeps the stdio transport alive and prevents the ~4 min timeout. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

During cancellation teardown, the inner task could raise a non-CancelledError exception (e.g. ValueError from model execution) which would mask the outer CancelledError. Catch Exception instead to ensure cancellation always propagates correctly. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add Hermes 4 (405B, 70B), Qwen 3.6 Plus, MiniMax M2.7, Qwen3 Coder, and Gemini 3.1 Flash Lite. Migrate primary aliases (qwen, minimax) to latest model versions. Total models: 35 → 41. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

gemini-code-assist

Code Review

This pull request updates the server to version 9.9.0 and increases the minimum Python requirement to 3.10. The changes include a major update to the OpenRouter model registry, adding several new frontier models (such as Claude 4.6, Gemini 3.1, and GPT-5.4) and migrating generic aliases to point to these latest versions. Additionally, a new progress notification mechanism was implemented in server.py to prevent client timeouts during long-running tool executions. Review feedback identifies a typo in the minimax-m2.7 model name and points out that some tests and documentation still reference older model versions for the qwen and minimax aliases.

gemini-code-assist · 2026-04-14T21:15:23Z

+            ("qwen", "qwen/qwen3.5-397b-a17b"),
+            ("minimax", "minimax/minimax-m2.5"),


These test cases for qwen and minimax aliases seem to be asserting the old mappings. According to the changes in openrouter_models.json and the PR description, these aliases should point to the newer model versions.

qwen should resolve to qwen/qwen3.6-plus.

minimax should resolve to minimax/minimax-m2.7.

Please update these assertions to reflect the alias migrations. Note that this suggestion assumes the model_name for minimax-m2.7 is corrected to use a dot instead of a dash, as suggested in another comment.

Suggested change

("qwen", "qwen/qwen3.5-397b-a17b"),

("minimax", "minimax/minimax-m2.5"),

("qwen", "qwen/qwen3.6-plus"),

("minimax", "minimax/minimax-m2.7"),

gemini-code-assist · 2026-04-14T21:15:23Z

+      "intelligence_score": 17
+    },
+    {
+      "model_name": "minimax/minimax-m2-7",


There appears to be a typo in the model_name. Based on the aliases (minimax-m2.7, m2.7) and the description, it should likely be minimax/minimax-m2.7 with a dot instead of a dash.

Suggested change

"model_name": "minimax/minimax-m2-7",

"model_name": "minimax/minimax-m2.7",

gemini-code-assist · 2026-04-14T21:15:23Z

+| `qwen`, `qwen3.5` | `qwen/qwen3.5-397b-a17b` | Frontier 397B MoE reasoning model (262K context) |
+| `minimax`, `m2.5` | `minimax/minimax-m2.5` | SWE-Bench 80.2%, agent-optimized (197K context) |


The documentation for the qwen and minimax aliases seems to be out of sync with the changes in openrouter_models.json.

The qwen alias now points to qwen/qwen3.6-plus, but the table here lists it with qwen/qwen3.5-397b-a17b.

The minimax alias now points to minimax/minimax-m2.7, but the table lists it with minimax/minimax-m2.5.

The PR description mentions these migrations, so the documentation should be updated to reflect the new state.

Suggested change

| `qwen`, `qwen3.5` | `qwen/qwen3.5-397b-a17b` | Frontier 397B MoE reasoning model (262K context) |

| `minimax`, `m2.5` | `minimax/minimax-m2.5` | SWE-Bench 80.2%, agent-optimized (197K context) |

| `qwen` | `qwen/qwen3.6-plus` | Alibaba's latest frontier model (1M context). `qwen3.5` -> 3.5 |

| `minimax` | `minimax/minimax-m2.7` | 1T param model matching GPT-5.3-Codex. `m2.5` -> 2.5 |

- Fix minimax-m2-7 typo → minimax-m2.7 (dot not dash, consistent with m2.5) - Update test assertions for migrated qwen/minimax aliases - Update docs alias table with new model mappings and new models Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 06127ed2ce

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-04-14T21:19:13Z

+            ("qwen", "qwen/qwen3.5-397b-a17b"),
+            ("minimax", "minimax/minimax-m2.5"),


Align alias-resolution test with migrated qwen/minimax defaults

These assertions still expect qwen and minimax to resolve to qwen/qwen3.5-397b-a17b and minimax/minimax-m2.5, but the updated registry now maps those aliases to qwen/qwen3.6-plus and minimax/minimax-m2-7 (conf/openrouter_models.json). As written, test_alias_resolution will fail once dependencies are installed, and it also blocks the intended alias migration from being validated.

Useful? React with 👍 / 👎.

timeleft-- and others added 13 commits March 6, 2026 07:29

merge: Resolve config.py version conflict with upstream main

8edb2f7

Keep version 9.9.0 from our feature branch (upstream synced to 9.8.2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

fix: Address review feedback on model configs

830f732

- gpt-5.3-codex: supports_temperature true→false (contradicts fixed constraint) - minimax-m2.5: max_output_tokens 196608→32768 (practical default; API allows up to 196K) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

style: Run black on pre-existing unformatted files

e96c33c

10 files were failing the black --check CI step before this PR. None are files modified by this PR — fixing them here to unblock CI. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

feat: Migrate codex alias to latest (gpt-5.3-codex)

35da1ca

- codex → openai/gpt-5.3-codex (was gpt-5-codex) - Old gpt-5-codex keeps codex-5.0 for backward compat Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/feat/update-openrouter-models-20…

30525fa

…26-03'

fix: log suppressed task exceptions during cancellation teardown

4ff6fc4

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

gemini-code-assist bot reviewed Apr 14, 2026

View reviewed changes

chatgpt-codex-connector bot reviewed Apr 14, 2026

View reviewed changes

timeleft-- mentioned this pull request Apr 14, 2026

feat: Add 9 new models, remove 2 obsolete, update aliases #411

Closed

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Add 6 new OpenRouter models (April 2026 refresh)#430

feat: Add 6 new OpenRouter models (April 2026 refresh)#430
timeleft-- wants to merge 14 commits intoBeehiveInnovations:mainfrom
MachineWisdomAI:feat/update-openrouter-models-2026-04

timeleft-- commented Apr 14, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Apr 14, 2026

Uh oh!

gemini-code-assist bot Apr 14, 2026

Uh oh!

gemini-code-assist bot Apr 14, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Apr 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		("qwen", "qwen/qwen3.5-397b-a17b"),
		("minimax", "minimax/minimax-m2.5"),

	"model_name": "minimax/minimax-m2-7",
	"model_name": "minimax/minimax-m2.7",

		\| `qwen`, `qwen3.5` \| `qwen/qwen3.5-397b-a17b` \| Frontier 397B MoE reasoning model (262K context) \|
		\| `minimax`, `m2.5` \| `minimax/minimax-m2.5` \| SWE-Bench 80.2%, agent-optimized (197K context) \|

Uh oh!

Conversation

timeleft-- commented Apr 14, 2026

Summary

Test plan

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant