July 2026

​

Signal and managed agents are generally available: Continuous monitoring and automated fix proposals

• Bring your own Anthropic key — mark an Anthropic provider as eligible for agents, then use it on individual presets, on Signal, or as the space-wide default. Enterprises that require customer-owned keys only can request that Arize disable the Arize-managed integration.
• Custom roles for managed agents — grant MANAGED_AGENT_READ, MANAGED_AGENT_CREATE, MANAGED_AGENT_UPDATE, and MANAGED_AGENT_DELETE independently to separate who can view agent runs from who can start them.
• Slack for managed agents — connect Slack and bind managed agents to the channels where your team already works.
• Skills for Agent-as-a-Judge — attach account-level skills to Agent-as-a-Judge evaluators.

​

Bind one service key across multiple organizations and spaces

ax api-keys create-service-key \
  --name "CI bot" \
  --assignments '[
    {"org_id":"<org-id-1>","role":"READ_ONLY","spaces":[{"space":"prod","role":"MEMBER"},{"space":"staging"}]},
    {"org_id":"<org-id-2>","spaces":[{"space":"sandbox","role":{"type":"CUSTOM","id":"<role-id>"}}]}
  ]'

​

Score traces with your own remote evaluators on Enterprise accounts

• Your own systems and dependencies — the endpoint runs in your environment, so it can call your databases, internal APIs, and any library live, including packages that are not available to code evaluations in Arize AX.
• Proprietary and non-LLM scorers — return a structured score or label from rule-based checks, machine learning models, or an internal gateway, with no need to conform to a supported LLM provider format.
• External eval services — wrap a third-party scorer and still get Arize AX orchestration and analytics around the results.
• Compliance and data residency — evaluation runs on your own infrastructure, and any auth tokens or secret headers you configure are encrypted at rest and masked in the UI.

@app.post("/evaluate")
def evaluate():
    fields = (request.get_json(silent=True) or {}).get("input") or {}
    return {"label": "ok", "score": 1, "explanation": "stub"}, 200

app.post("/evaluate", (req, res) => {
  const fields = (req.body && req.body.input) || {};
  res.status(200).json({ label: "ok", score: 1, explanation: "stub" });
});

​

Find your data region’s OTLP and Flight endpoints in settings

• Connection endpoints — copy your region’s OTLP HTTP, OTLP gRPC, and Flight endpoints straight from the Data Region section.
• Personal preferences — switch between light, dark, and system themes, set your display time zone independently of the active time range, and change your password.

​

Run Claude Opus 5 in the Playground, Prompt Hub, and evaluators

• Anthropic — claude-opus-5
• AWS Bedrock — us.anthropic.claude-opus-5 and global.anthropic.claude-opus-5
• Vertex AI — claude-opus-5@default

​

Fixes and improvements

• Fix Sort the ML Models list page by name, volume, or created date again.

• Fix Run models on custom OpenAI-compatible endpoints that require max_completion_tokens in the Playground, which now sends the token limit these endpoints expect.

• Fix Align an evaluator in Alyx reliably, with the evaluator loaded into the Playground before alignment begins.
• Fix View the evaluator list in Alyx for datasets whose evaluation tasks include a code-based evaluator.

• Fix Read experiment evaluation and annotation scores accurately in the comparison view, with non-finite values now rejected at ingest.

• Fix Manage custom OpenAI-compatible model endpoints through the /v2/integrations REST API, where they now return with the CUSTOM type.
• Improvement Work with experiments that have no associated dataset in the JavaScript and TypeScript SDK, which now accepts null dataset and version IDs.
• Improvement Grant or revoke API key creation for SAML-mapped users with a new developer-access control on each role mapping.

• Improvement Narrow the user activity view with new exclude filters on source and kind, not just match filters.

​

Chart dataset and experiment distributions on your dashboards

​

Run Gemini 3.6 Flash and Gemini 3.5 Flash-lite in the Playground and evaluators

​

Manage annotation configs, dataset examples, and restrictions from the CLI

# Update the allowed values and optimization direction of a categorical annotation config
ax annotation-configs update categorical "response-quality" \
  --value good --value bad \
  --optimization-direction MAXIMIZE

# Delete specific examples from a dataset version
ax datasets delete-examples "my-dataset" \
  --version-id 0a1b2c3d-4e5f-6a7b-8c9d-0e1f2a3b4c5d \
  --example-ids '["8f14e45f-ea1c-4b3a-9d21-2b0c0a1f9e77"]'

# List the project restrictions in your space
ax resource-restrictions list --resource-type PROJECT

​

Review Signal issues across your whole agent fleet

​

Connect AWS Bedrock, Vertex AI, NVIDIA NIM, and custom endpoints through the REST API

curl -X POST https://api.arize.com/v2/integrations \
  -H "Authorization: Bearer $ARIZE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "LLM",
    "name": "vertex-ai-prod",
    "config": {
      "provider": "VERTEX_AI",
      "project_id": "my-gcp-project",
      "location": "us-central1",
      "project_access_label": "arize-vertex-access"
    }
  }'

ax ai-integrations create \
  --name "vertex-ai-prod" \
  --provider VERTEX_AI \
  --provider-metadata '{"project_id": "my-gcp-project", "location": "us-central1", "project_access_label": "arize-vertex-access"}'

​

Save and reuse custom dataset views

​

Fixes and improvements

• Fix Use Anthropic Claude models through Vertex AI again, with request headers and parameters handled correctly for Vertex’s API.
• Fix Run GPT-5 and o-series models in the Playground on Azure, which now sends the token limit these models require.
• Improvement Evaluate longer sessions and traces with gpt-4.1, now supported up to a 1M-character context.

• Fix Reference dataset columns by name in chained evaluators, including columns whose names match experiment outputs such as input or output.
• Improvement Create evaluators with the scope preset to match the template you start from, so Session and Trajectory templates no longer default to Span.
• Fix Create and fetch template evaluators through the SDK and CLI without deserialization errors from the new use_structured_output field.

• Fix Track LLM cost on spans imported through file upload, which now preserves the cost attributes on ingest.
• Improvement See tool-call arguments in the Input panel for tool spans sent with OpenTelemetry GenAI conventions.
• Fix Share a direct link to a specific session that opens with full context on the dashboard.
• Fix Jump to the latest available data reliably in an empty project view, without a previous project’s time range carrying over.

• Fix See correct timestamps at fine granularities in half-hour-offset time zones.

• Improvement Pass naive timestamps to any AX CLI command and have them interpreted as UTC.

• Fix Keep shared service API keys active after removing the user who created them.

​

See all experiments in one place, grouped your way

​

REST API v2 now uses consistent casing and schema names

• Enum recasing — values such as admin, openAI, and template_evaluation are now ADMIN, OPEN_AI, and TEMPLATE_EVALUATION.
• Verb-first schemas — response types such as IntegrationListResponse and SpanDeleteResponse are now ListIntegrationsResponse and DeleteSpansResponse.

# Fetch the latest OpenAPI 3.1 spec, then point your client generator at openapi.yaml
curl -fsSL https://api.arize.com/v2/spec.yaml -o openapi.yaml

​

Model registry lists only actively supported models

• End-of-life removals — retired OpenAI, Anthropic, Vertex AI, AWS Bedrock, and NVIDIA NIM snapshots, including gpt-4-0314 and the legacy Claude 3, 3.5, and 3.7 versions.
• Gemini 2.0 deprecation — gemini-2.0-flash, gemini-2.0-flash-001, and gemini-2.0-flash-lite are removed, and saved prompts that referenced them fall back gracefully.

​

Agent Playground and Remote Agents are generally available

​

Update existing dataset examples in place with the Python SDK

import os

from arize import ArizeClient

client = ArizeClient(api_key=os.environ["ARIZE_API_KEY"])  # read the key from the environment

updated_examples = [
    {
        "id": "your-example-id",  # the example to update; IDs that don't exist are ignored
        "expected_output": "Paris (updated)",
        "eval.Correctness.label": "correct",
    },
]

result = client.datasets.update_examples(
    dataset="dataset-name-or-id",
    space="your-space-name-or-id",  # required when passing a dataset name
    examples=updated_examples,
    # new_version="v2-fixed-answers",  # optional: capture the update as a new version
)

print(result.dataset_version_id, result.example_ids)

​

Attach multiple session and trace evaluators to a single task

​

Self-serve 13-month data retention on Pro plans

​

New OpenAI GPT-5.6 models in the Playground and evaluators

• Model options — gpt-5.6-sol (frontier), gpt-5.6-terra (balanced), and gpt-5.6-luna (efficient).
• Reasoning effort — choose from none, low, medium, high, xhigh, and the new max level.

​

Bring your own Anthropic key for managed agents

​

Fixes and improvements

• Fix Track accurate cost and token totals across agent and chain traces that report SDK rollups.
• Improvement Search on span start time, end time, and latency in the attributes panel, and include them in span exports.

• Fix Get completed results from evaluators you apply directly in the Playground.

• Fix Select Gemini 2.5 and newer models in the Playground model picker.
• Fix Track accurate trace cost for Vertex AI models.
• Fix Run non-pro GPT-5 models against custom and OpenAI-compatible endpoints that implement only Chat Completions.

• Fix Filter distribution widgets and monitors by the values you set.

• Fix Scope annotation configurations to the specific trace you apply them to.

• Improvement Submit dataset example annotations once at least one row is labeled.
• Improvement Tell Alyx how many spans to run a task on and whether to skip already-evaluated spans.

​

Enter spaces with any resource role you have been granted

• Space entry gate — holding any binding in the space lineage, or legacy membership, is enough to enter the space.
• Full coverage — restores both the in-app space view and the GET /v2/spaces and GET /v2/spaces/{id} endpoints for custom-role users.

​

Control Arize AX’s appearance with a System theme and refreshed light mode

• System theme — choose System, Light, or Dark; System follows your OS appearance and switches live as it changes, with no reload.
• Refreshed light mode — a warm-neutral palette with layered elevation and vivid success, warning, danger, and info status colors.
• Accessibility — links, severe text, and input placeholders now meet WCAG AA contrast.

​

Fixes and improvements

• Improvement Alyx can now order aggregated results and return the top-N groups by a column you choose, so it can answer ranking questions such as which models drive the most cost.

• Improvement Code evaluator test runs now send only the dataset columns you have mapped to the evaluator’s arguments, so the test payload matches what the evaluator receives.

• Fix You can now drag to select text in the prompt variable editor.

• Improvement The trace slideover now hides the Evaluations and Annotations tabs by default and omits internal-only metadata from evaluation tables; re-enable the tabs anytime from Customize Tabs.
• Fix Trace field values that contain non-ASCII, multibyte characters are now truncated on character boundaries, so international text stays intact and valid.

• Fix You can now filter pivot table widgets on null or empty values and get the matching rows.

• Fix Webhook endpoints now accept requests that include a trailing query string, so integrations that append one deliver successfully.

​

Evaluate agents with new prebuilt trajectory and session evaluators

• New agent and session evaluators — Goal Completion, Path Efficiency, Reasoning Coherence, Session Resolution, Topic Coherence, Session Frustration, and Session Completion.
• Organized by workflow — templates are grouped into Response Quality, Code Quality, Trajectory, Session, RAG, and Security so you can find the right evaluator faster.
• Scope at a glance — each template shows whether it runs on a span, trace, or session.
• Security evaluators for everyone — the Security template group is now available to all users.

​

Capture more LLM and tool detail with new OpenInference span attributes

• llm.system — the system or provider that served the LLM call.
• llm.finish_reason — why the model stopped generating.
• tool.id — the tool-call identifier, now available as a top-level attribute.
• embedding.invocation_parameters — the parameters used for embedding calls.

​

Run prompts and evaluations on Claude Sonnet 5

• Available everywhere you pick a model — select Claude Sonnet 5 in the Prompt Playground and when configuring evaluators.
• Automatic cost tracking — input, output, and cache token costs are recorded for every call, so spend shows up in your usage and evaluation costs.
• Adaptive reasoning — like other latest Claude models, Sonnet 5 manages its own reasoning effort, so the temperature, top-p, and top-k controls no longer apply.

​

Map evaluator variables to dataset columns in the Prompt Playground

• Per-instance variable mappings — a Mappings control on each evaluator instance maps template variables such as {{input}} to dataset columns such as question, and warns you when a mapping is missing.
• Automatic mapping with Alyx — the Align Eval flow detects column mismatches and applies the correct mappings before running.
• Preview with resolved values — the preview table shows each variable → column mapping alongside the resolved values.

​

Launch prebuilt agents with a guided setup wizard

• Ready-made agents — start from templates like SRE/On-Call, Investigate Failing Traces, Fix a Bug, Incident Commander, and Cost.
• Guided Connect, Configure, and Review — connect skills, point the agent at a project or repository, and launch it as a session or a recurring automation.

​

Analyze experiment trends with the new Experiment Analysis view

• Trend chart — track how scores and metrics change across experiment runs over time.
• Scoreboard — compare the key metrics for your selected experiments side by side.

​

Fixes and improvements

• Improvement You can now start an evaluator run on a dataset directly from the slideover in a single step.

• Improvement Your span Input/Output and Attributes display format (Pretty, Raw, JSON) now persists across sessions.
• Improvement The Projects flyout now lists recently viewed projects with a 7-day volume sparkline, and prompt tools appear as top-level navigation items.
• Improvement Added a Clear Filters action to the tracing table empty state so you can recover from a filtered-empty view in one click.

• Improvement Automation detail slide-overs now show the agent’s investigation report, with clickable references to the traces it examined.