August 2026 - Arize AX Docs

List traces for a project with the Python SDK

August 18, 2026 New SDKs and REST APIs The Python SDK now includes client.traces.list(), so you can retrieve a project’s traces in code, each returned with its full list of spans and roll-up metadata such as the root span ID and a spans-truncated flag. Pass a project by name (with a space) or by ID, and narrow the results with an optional time range, a filter expression, and a result limit; the server returns an opaque cursor for paging through large result sets. The filter uses the same expression syntax as client.spans.list() and selects any trace that contains at least one matching span. This is a beta endpoint and may change without notice. The code example below lists a project’s most recent error traces and reads the cursor for the next page:

import os

from arize import ArizeClient

client = ArizeClient(api_key=os.environ["ARIZE_API_KEY"])

# List the most recent traces for a project, keeping only those with an error span.
response = client.traces.list(
    project="my-project",            # project name (or pass an ID)
    space="my-space",                # optional when project is an ID
    filter="status_code = 'ERROR'",  # selects traces with at least one matching span
    limit=50,                        # optional, defaults to 50
)

for trace in response.traces:
    # Each trace carries its full list of spans plus roll-up metadata.
    print(trace.trace_id, trace.root_span_id, len(trace.spans))

# Page through larger result sets with the opaque cursor the server returns.
next_cursor = response.pagination.next_cursor

Run Gemini 3.7 Flash in the Playground and evaluators

August 18, 2026 New Models and Integrations Select gemini-3.7-flash from the model picker to prototype prompts and score traces with Google’s latest Flash model. It is available through both the direct Gemini and Vertex AI providers, accepts text, image, audio, and video input, and exposes a thinking level control you can set to low, medium, or high. Learn more about model integrations.

Fixes and improvements

August 14–19, 2026 Evaluators

Improvement Add a remote evaluator to a project from the tasks menu in its tracing view, so you can attach your hosted scoring endpoint without leaving the traces page.
Improvement Delete an online evaluation task, which also removes the experiment runs associated with it.
Fix See accurate failure totals in online eval runs, with rows that end in an exception now counted alongside failed rows.
Fix Run online evaluations across all your datasources: scheduling now rotates through every datasource each cycle so those beyond the current task-slot capacity still get their turn.

Playground

Fix Keep image variables listed in the Input Variables column when a message’s only content is the image reference.

Alyx

Improvement Choose the gpt-5.6-terra model for Alyx.

View uploaded images and audio throughout Arize AX

August 11, 2026 New Tracing and Sessions Media you upload to Arize AX now resolves to secure, viewable content wherever it appears, so images and audio render in place of raw references while you work. Uploaded media displays in trace and span views, datasets, experiment comparisons, the prompt playground, and annotation queues, and it renders consistently whether your data lives in Google Cloud, AWS, or Azure. When a file cannot be shown, the view states why, distinguishing media that is still processing, an upload that failed, and a reference that was not found. Media upload is now enabled in the US, EU, and CA production regions, and media file storage is counted in the Plans & Usage storage total.

Create standalone experiments from the CLI and review their runs

August 11, 2026 Improvement Datasets and Experiments You can now create an experiment that is not attached to any dataset directly from the CLI and review its runs, metrics, and evaluation scores in a dedicated runs table in the UI. Pass --space in place of --dataset to ax experiments create to record runs produced outside a dataset, then open the experiment to analyze results without dataset-specific columns. This extends dataset-less experiments, previously available in the SDKs and REST API, to the CLI and the experiment views. The code example below creates a standalone experiment scoped to a space:

# --space is required when --dataset is omitted, and says where the standalone experiment lives.
ax experiments create --name "My Experiment" --space "your-space-name-or-id" --file runs.csv

Visit the experiments CLI documentation to learn more.

Route OTLP traces to a project with a request header

August 11, 2026 New Tracing and Sessions Generic OTLP/HTTP clients can now send traces to a specific project by setting an x-project-name request header, so tools that cannot set an OpenInference project resource attribute (for example, no-code platforms such as n8n) still land in the right project. A project named in the span or resource payload still takes precedence over the header, and a request that carries no routing information now returns a clear InvalidArgument error instead of a misleading permission failure. The code example below adds the routing header to an OTLP/HTTP trace exporter:

new OTLPTraceExporter({
  url: "https://otlp.arize.com/v1/traces",
  headers: {
    "arize-space-id": process.env.ARIZE_SPACE_ID!,
    "arize-api-key": process.env.ARIZE_API_KEY!,
    "x-project-name": "my-project", // routes spans that carry no project of their own
  },
});

Learn more about the OpenTelemetry exporter.

Select time ranges beyond your data retention window

August 10, 2026 Improvement Tracing and Sessions Time range presets and custom dates that reach past your account’s retention window are now selectable across the Traces, Spans, and Sessions tables instead of being disabled. Rows within your retention window render normally, while the portion of the range that falls outside it appears as a blurred section with an upgrade prompt, so you can see exactly where your accessible data ends. This gating now also applies to the Sessions tab, which previously had none.

Evaluate images with LLM-as-a-judge templates

August 8, 2026 New Evaluators LLM-as-a-judge evaluator templates can now reference images, so you can build vision evaluations that score the picture a span produced or received rather than a text stand-in. Reference an image column in a template and Arize AX sends the image to the judge model, whether it is provided as a public URL, inline base64, a cloud-storage object, or an uploaded Arize media reference. You can add an image reference from the template column picker, preview it before saving, and test the evaluator against real spans before you run it. Visit the LLM as a Judge documentation to learn more.

Let Alyx configure filters for trace and session evaluations

August 8, 2026 New Alyx Alyx can now set up the multi-span query filters and per-variable subquery mappings that trace-level and session-level evaluation tasks require, so you no longer configure them by hand. When you ask Alyx to build a trace or session evaluation, it reads your open evaluator or task form, proposes the filters and variable mappings with a data preview, and asks before changing filters you have already set. It also flags queries that would match no traces, preventing evaluation tasks that run but score nothing. Learn more about Alyx.

Run Mistral models through AWS Bedrock

August 7, 2026 New Models and Integrations Mistral models hosted on AWS Bedrock are now available in the model picker for the prompt playground and evaluations, grouped under Mistral AI, so you can run prompts and LLM-as-a-judge evaluations on Mistral through your Bedrock integration. The lineup adds:

Mistral Large 3, Devstral 2, and Magistral Small
Ministral 3 in 14B, 8B, and 3B sizes
Pixtral Large, for vision workloads

For more information, refer to the Amazon Bedrock documentation.

Choose which trace metrics appear in the trace slideover header

August 6, 2026 New Tracing and Sessions Metrics you send as metric.* span attributes are now ingested as first-class values and can be pinned to the header of the trace slideover. Open the metrics dropdown in the trace slideover header to choose which metrics display as name and value tokens beside the trace, so the numbers you track most are visible without opening individual span attributes. The dropdown appears only for traces that carry metrics.

Fixes and improvements

August 6–12, 2026 SDKs and REST APIs

Improvement Pin a task’s evaluator to a specific version through the v2 REST API by setting evaluator_version_id on POST /v2/tasks or PATCH /v2/tasks/{task_id}.
Improvement Scope a triggered experiment run to a filtered subset of dataset examples by passing query_filter on POST /v2/tasks/{task_id}/trigger, combinable with example_ids.

Models and Integrations

Improvement Point an Anthropic integration at a custom API base URL from the integration configuration or the v2 REST API, so you can route requests through a proxy or compatible endpoint.

Tracing and Sessions

Fix Filter by trace-level and session-level evaluations from the tracing filter dropdown, including in projects that have only trace or session evals.
Fix See the current turn in the Span Input column, which now reads the latest user message in a multi-turn conversation instead of the first.

Playground

Fix Keep the Agent Playground output visible while editing request JSON, which no longer crashes on partially typed input.

Webhooks and Events

Fix Add or change a webhook’s authentication token when editing an existing webhook.

Dashboards and Visualizations

Improvement Plot every value of a dimension on a line chart with a select-all live-update toggle, so new values appear automatically without editing the widget.

Alyx

Fix Keep the Alyx dock open while editing a dashboard widget.

Agents

Improvement Sort detected issues in Signal by severity or last seen.

CLI Commands

Fix Use the documented US, EU, and CA region aliases with ax profiles create and ax profiles update.

Create experiments without a dataset

August 5, 2026 New Datasets and Experiments You can now create an experiment scoped to a space instead of a dataset, so runs and results you produced elsewhere can be recorded in Arize AX even when there is no dataset to attach them to. In the Python SDK, pass space in place of dataset to experiments.create(); through the REST API, send space_id instead of dataset_id on POST /v2/experiments and filter dataset-less experiments with the new space_id query parameter on GET /v2/experiments. This extends dataset-less experiment support, previously available in the JavaScript and TypeScript SDK, to Python and the REST API. The code example below creates a standalone experiment in a space:

from arize.experiments import ExperimentTaskFieldNames

# Pass `space` instead of `dataset` to create a standalone experiment
# that is not tied to any dataset.
experiment = client.experiments.create(
    name="my-standalone-experiment",
    space="your-space-name-or-id",
    experiment_runs=[{"output": "the task output"}],
    task_fields=ExperimentTaskFieldNames(output="output"),  # example_id is optional for standalone runs
)

Visit the experiments client documentation to learn more.

Delete spans faster by scoping to a time window

August 5, 2026 Improvement SDKs and REST APIs Deleting spans by ID through DELETE /v2/spans now accepts optional start_time and end_time bounds, so a delete searches only the interval when the spans were ingested instead of the full lookback window. Each bound is independent: omit start_time and it defaults to the start of the lookback window, omit end_time and it defaults to now. The parameters are available across the Python, Go, and TypeScript SDKs. The code example below deletes two spans, scoping the search to a single day:

from datetime import datetime, timezone

# Scope the delete to the day the spans were ingested.
client.spans.delete(
    project="your-project-name-or-id",
    span_ids=["span-id-1", "span-id-2"],
    start_time=datetime(2026, 8, 4, tzinfo=timezone.utc),  # inclusive lower bound
    end_time=datetime(2026, 8, 5, tzinfo=timezone.utc),    # exclusive upper bound
)

Visit the spans client documentation to learn more.

Write new API keys directly to a .env file

August 3, 2026 New CLI Commands The ax api-keys create and ax api-keys create-service-key commands now take an --env-file flag that writes the generated key to ARIZE_API_KEY in a dotenv or .envrc file, so the new credential is written into your local environment and the raw key is never printed to your terminal. The write preserves the file’s existing formatting, and if it fails, the new key is automatically revoked so you are never left with an unusable credential. The code example below creates a key and writes it into a local dotenv file:

# Write the new key as ARIZE_API_KEY into .env.local.
# If the file write fails, the key is automatically revoked.
ax api-keys create --name "Local development" --env-file .env.local

Visit the api-keys CLI documentation to learn more.

Pin evaluators to a specific version, now on every account

July 31, 2026 New Evaluators Evaluator version pinning is now available on every account. Pin an evaluator task to a specific published version and the task keeps running that version when newer versions are published, so a task you have tuned and validated does not change when someone publishes an update. Choose “Latest” when you would rather a task always pick up the newest version. You select the version from the evaluator’s version dropdown wherever you configure the task. Learn more about creating evaluators.

Fixes and improvements

August 3–5, 2026 Evaluators

Improvement Skip a mapping step when building template evals on a project: the {context} variable now maps to the input column automatically.

Tracing and Sessions

Fix Open span details for backfilled spans whose event time differs significantly from their ingestion time.

Dashboards and Visualizations

Fix See dashboard widgets and Actual Class filter values scoped to the date range you selected, rather than the full account lookback.
Fix See correct monthly buckets in metrics and dashboards when your time zone has a non-zero UTC offset.

Webhooks and Events

Improvement Reuse the name of a webhook destination you previously deleted when creating a new one.

Custom Metrics

Fix Filter custom metrics with a contains condition on string-list fields such as agent tool calls.

SDKs and REST APIs

Fix Retrieve task runs from the Python SDK reliably as the API adds new response fields.

​List traces for a project with the Python SDK

​Run Gemini 3.7 Flash in the Playground and evaluators

​Fixes and improvements

​View uploaded images and audio throughout Arize AX

​Create standalone experiments from the CLI and review their runs

​Route OTLP traces to a project with a request header

​Select time ranges beyond your data retention window

​Evaluate images with LLM-as-a-judge templates

​Let Alyx configure filters for trace and session evaluations

​Run Mistral models through AWS Bedrock

​Choose which trace metrics appear in the trace slideover header

​Fixes and improvements

​Create experiments without a dataset

​Delete spans faster by scoping to a time window

​Write new API keys directly to a .env file

​Pin evaluators to a specific version, now on every account

​Fixes and improvements

List traces for a project with the Python SDK

Run Gemini 3.7 Flash in the Playground and evaluators

Fixes and improvements

View uploaded images and audio throughout Arize AX

Create standalone experiments from the CLI and review their runs

Route OTLP traces to a project with a request header

Select time ranges beyond your data retention window

Evaluate images with LLM-as-a-judge templates

Let Alyx configure filters for trace and session evaluations

Run Mistral models through AWS Bedrock

Choose which trace metrics appear in the trace slideover header

Fixes and improvements

Create experiments without a dataset

Delete spans faster by scoping to a time window

Write new API keys directly to a .env file

Pin evaluators to a specific version, now on every account

Fixes and improvements